GitHub Agentic Workflows | Blog

Agent of the Day – October 30, 2026

Fri, 30 Oct 2026 00:00:00 GMT

Every backlog needs someone with an appetite. Today’s spotlight goes to a workflow that wakes up every half hour, sniffs around for open issues, and quietly hands them to the Copilot coding agent.

Agent of the Day: Issue Monster

Issue Monster calls itself the Cookie Monster of issues. It runs on a 30-minute schedule and can also be dispatched by hand. It is deliberately polite about it: it skips a run when five or more Copilot draft pull requests are already open, and when there are no open issues there is nothing to eat.

What the live runs show

We pulled the last three days of runs with agentic-workflows logs. Five Issue Monster runs came back, all from this morning (UTC), four of them successful:

Run	Result	Duration	Tokens	AI credits
41065024612	success	8.8 min	73,963	0.56
41061735192	success	8.6 min	114,945	0.84
41057495921	failure	7.5 min	45,623	0.52
41054891396	success	8.8 min	75,756	0.60
41049553200	success	9.1 min	75,048	0.60

Each successful run posted three comments, so twelve safe-output comments in total. Some issues came up more than once, which is a nice reminder that triage is a conversation rather than a one-shot: #73041 was commented on in four runs, #73122 in two, and #72996 and #72835 in two each. Single visits went to #72977 and #72983.

The one that tripped

Not every run is a victory lap. An audit of run 41057495921 flagged a critical finding: the job could not resolve a pinned actions/checkout commit, so the run ended in failure. That is a platform-side resolution error rather than a logic bug, and the very next scheduled run was green again. The same audit also noted, at low severity, that this issue-response task might run on a cheaper model. Worth a look for anyone tuning costs.

Why it matters

Across the five runs the workflow used roughly 385,000 tokens and about 3.1 AI credits. That’s a small price for keeping the flow of issues to Copilot steady, visible, and auditable.

Try it yourself

Browse the workflow source, run gh aw audit on your own runs, and steal the pattern at github/gh-aw.

Agent of the Day – October 29, 2026

Thu, 29 Oct 2026 00:00:00 GMT

Code review is a craft, and a tiring one. Today’s spotlight goes to a workflow that shows up the moment a pull request is marked ready and gives it a careful second pair of eyes.

Agent of the Day: PR Code Quality Reviewer

PR Code Quality Reviewer covers bugs, performance, style, naming and best practices in a single pass. Its description says it consolidates the older Grumpy Code Reviewer and PR Nitpick Reviewer. It triggers when a pull request moves to ready_for_review (docs-only and changeset-only changes are skipped), or on demand through the /review slash command.

What the live runs show

We pulled the last three days of runs with agentic-workflows logs. Five runs came back, all successful:

Run	Pull request	Duration	Tokens	AI credits
40871527808	#72940	9.4 min	17,774	2.32
40871500692	#72941	10.6 min	28,865	3.04
40869752898	#72938	10.6 min	33,268	4.69
40869617197	#72932	10.4 min	19,536	2.96
40868959452	#72958	8.3 min	3,738	1.79

The summary reports 49.3 minutes of total duration, about 103,000 tokens, roughly 14.8 AI credits and 16 safe-output items. Every run took two turns, and there were no errors, missing-tool requests or missing-data requests. The model recorded in the runs is gpt-5.6-luna.

The audit view

We ran audit on run 40871500692. It compared the run against an earlier baseline, 40869752898, matched on event, task domain and resource profile, and labelled the pair “stable”: same turn count, same write-capable posture, zero blocked requests.

The audit also raised one flag worth being honest about. It classified the run as a heavy profile for a code-fix task: two turns, nine tool types and six write actions over about ten and a half minutes. Its recommendation is to compare against similar runs and trim unneeded tools or writes. The spread in token use above (3,738 to 33,268) suggests review depth tracks the size of the pull request, which is plausible but is our reading, not something the logs state.

Why we like it

It is predictable. Five reviews, five successes, a consistent shape. For a reviewer, being boring in the right way is a feature.

Try it

Browse the workflow source, then adapt it for your own repository at github/gh-aw.

Agent of the Day – October 28, 2026

Wed, 28 Oct 2026 00:00:00 GMT

Backlogs rarely die in a single heroic push. They shrink when someone keeps nibbling at them. Today’s spotlight goes to the workflow that does the nibbling, and does it on a leash.

Agent of the Day: Issue Monster

Issue Monster picks up open issues and leaves a comment on each one it handles. The live logs show it running on the Pi engine with the copilot/gpt-5 model, and each run is a compact affair: roughly eight to ten minutes, around 60,000 to 83,000 tokens, and three comments.

What the live runs show

We pulled the most recent runs with agentic-workflows logs. The window we collected spans just over two hours this morning (UTC):

Run	Started (UTC)	Duration	Outcome
40687259037	12:22	9.5 min	Success, 3 comments
40689976483	12:42	8.1 min	Success, 3 comments
40694785400	13:15	9.3 min	Success, 3 comments
40698326997	13:38	9.2 min	Success, 3 comments
40704218234	14:15	6.6 min	Failure (driver exit)

The summary reports 4 of 5 runs successful, 12 safe-output items, about 284,000 tokens and roughly 2.5 AI credits in total. The agent made zero missing-tool or missing-data requests.

Look at where the comments landed. Issue #72589 received a comment in three consecutive runs, and #72330 in three of the four. Others included #72119, #72479, #70975 and #68171. Whether that repetition is healthy follow-up or a sign the workflow keeps circling the same threads is a fair question for its maintainers, and we are not going to pretend the logs answer it.

The failed run, honestly

The 14:15 run ended in a driver_exit failure at the “Execute Pi CLI” step, with zero tokens recorded. The conclusion job’s failure handler fired as designed. This is the kind of detail a spotlight should include: an agent that runs every half hour will sometimes stumble, and what matters is that the failure is visible and classified rather than silent.

A caveat on evidence: our attempt to run a structured audit on run 40698326997 timed out during report generation, so the figures above come from the logs output alone.

Why it earns the spotlight

Small, regular, bounded runs make an agent easy to reason about. Each run costs well under a credit and produces a handful of comments, all through safe outputs, so a bad day is cheap and easy to review.

Try it

Read the source, borrow the pattern, and build your own at github.com/github/gh-aw.

Agent of the Day – October 27, 2026

Tue, 27 Oct 2026 00:00:00 GMT

Documentation drifts. A flag gets renamed on Tuesday, a feature lands on Wednesday, and by Friday the docs describe a product that no longer exists. Today’s spotlight goes to the workflow that treats that drift as a daily chore rather than a quarterly cleanup.

Agent of the Day: Daily Documentation Updater

The Daily Documentation Updater wakes up around 10:00 UTC every day, reads recent repository activity, and proposes documentation edits as a pull request. It runs on the Codex engine with read-only permissions on contents, issues and pull requests, and a restricted network allowlist (defaults and github). It cannot write to the repository itself. Its only way to change anything is the create-pull-request safe output, which opens a non-draft PR labeled documentation and automation, set to expire after one day.

What the live runs show

We pulled the last three days of runs with agentic-workflows logs:

Run	Date (UTC)	Duration	Outcome
40185702590	Oct 25	8.6 min	Opened PR #71867
40309934567	Oct 26	8.5 min	Succeeded, no PR
40490864186	Oct 27	9.6 min	Opened PR #72406

Across the three runs the log summary reports 3 of 3 successes, zero errors, zero missing tools, and zero agent logic failures. Total usage was roughly 3.1 million tokens, or about 527 AI credits, over 26.7 minutes of runtime.

The middle row is the one worth dwelling on. A run that finishes cleanly and opens nothing is a feature: when there is no documentation gap, the agent has no reason to manufacture one. Two out of three days produced a PR, and the day with nothing to say stayed quiet.

Why it earns the spotlight

Most documentation tooling checks that docs exist. This workflow checks that they are still true, and it does so on a schedule, so the cost of drift stays small. The change history is recent too: the latest commit touching the workflow, #72396, migrated it to lock-safe pull request outputs, which shows it is being maintained like any other piece of production code.

It is also a good template. The pattern is a schedule, minimal permissions, one narrow output, and a cap on daily spend (max-daily-ai-credits is set in the frontmatter). You can copy that skeleton for any recurring housekeeping task.

Try it

Browse the workflow source, then compile your own daily helper at github.com/github/gh-aw.

Before the Agent Starts: The Activation Gate

Tue, 27 Oct 2026 00:00:00 GMT

A scheduled workflow is ready to run. Its trigger matches, its authorization checks pass, and its prompt is compiled. But should it start another agent session right now?

That question can depend on more than one repository: an organization-wide budget, a workflow’s priority, or a centrally managed allowance. The activation gate adds a place to make that decision before the agent starts. In gh-aw configuration, the feature is called admission, enabled with on.admission: true.

gh-aw implements the workflow-side client and compiler wiring. A separately deployed admission authority owns policy, identity verification, and trusted usage accounting.

A decision before inference

For an opted-in workflow with a configured endpoint, the compiler-owned admission job runs before activation. It requests a GitHub Actions OIDC identity token for the authority’s audience, then posts the invocation identity and compiled workflow hash to /api/admission.

The authority verifies the identity and evaluates its policy. An approval supplies the primary agent’s maximum AI Credits, with optional independent allowances for threat detection and evals:

{
  "max_ai_credits": 1000,
  "max_ai_credits_threat_detection": 100,
  "max_ai_credits_evals": 50
}

These are illustrative allowances, not built-in tiers. Each supplied allowance replaces that component’s runtime limit, even if it is higher than the workflow’s configured limit. Omitted optional allowances leave the existing component limits unchanged; the three values are not a shared pool.

A valid approval supersedes the local daily-credit check for that activation. Any required shared-budget decision therefore belongs in the authority. An explicit denial stops activation before the agent runs.

Enable once, configure centrally

Add admission alongside a workflow’s existing triggers:

on:
  workflow_dispatch:
  admission: true

To opt in repository-wide instead, add "admission": true to the existing .github/workflows/aw.json object, preserving its other settings. A workflow-level false cannot override that repository opt-in.

Recompile affected workflows with gh aw compile and commit the source configuration and generated .lock.yml files. Admission jobs are omitted by default: setting an endpoint alone does not generate them.

The runtime configuration comes from organization or repository Actions variables:

Variable	Purpose
`GH_AW_DEFAULT_ADMISSION_URL`	Full HTTPS endpoint ending in `/api/admission`
`GH_AW_DEFAULT_ADMISSION_AUDIENCE`	OIDC audience; defaults to the endpoint URL
`GH_AW_DEFAULT_ADMISSION_ON_OUTAGE`	`ignore` by default, or `block`

Once the jobs are compiled in, changing these variables does not require recompilation. An empty endpoint skips admission and retains existing local checks. Repository variables take precedence over organization variables.

Authentication uses GitHub Actions OIDC, not a PAT or a provider credential. The feature grants id-token: write to its isolated admission and completion-callback jobs, not to the agent merely because admission is enabled. Hand-authored reusable-workflow callers must also permit that scope.

For managed-defaults commands and rollout details, see centralized cost controls.

Where CAO fits

For a CAO integration, this admission contract is the boundary between the executing workflow and centralized policy. A CAO deployment acting as the authority would verify workflow identity, apply its protected policy, and return allowances through the same protocol. The workflow does not need a CAO-specific credential or custom admission script.

That is an integration contract, not a claim that installing gh-aw deploys CAO or a hosted admission server. The authority must be provisioned and implement the admission service specification.

This cost-admission check is also distinct from the experimental CAO gateway trigger path. Gateway provenance and hash-bound admission validate incoming dispatches; activation admission decides whether a participating execution may start with a particular credit allowance.

Close the loop after execution

The same opt-in also generates conclusion_admission. After the ordinary conclusion job publishes the unified-session artifact, this isolated job sends bounded compact metadata and artifact IDs to /api/conclusion on the authority’s origin, using a fresh OIDC token.

The authority can use the callback to process completion metadata and retrieve the full session through authenticated GitHub artifact access. It is end-of-session delivery, not live streaming or billing settlement. Delivery is best-effort: a callback failure warns without changing completed agent or safe-output results.

A cooperative gate, not a hard cap

Outage behavior is an explicit availability choice. With the default ignore, an exhausted transient outage falls back to existing local controls. With block, it stops activation. Explicit denials and invalid approvals block under either policy.

Admission remains cooperative. Workflow editors can remove participation, repository-variable administrators can override organization defaults, and concurrent runs can overshoot a shared budget. Hard quotas, reservations, and mid-run cancellation require independently enforced backend controls.

The activation gate provides a focused extension point: make the start-time decision centrally, return explicit runtime allowances, and receive completion metadata afterward, while keeping policy-server responsibilities separate from workflow execution.

Agent of the Day – October 26, 2026

Mon, 26 Oct 2026 00:00:00 GMT

Every codebase has a file that started small and kept saying yes. Today’s spotlight goes to the workflow whose whole job is to notice.

Agent of the Day: Daily File Diet

Daily File Diet runs on a weekday schedule (around 13:00 UTC) and can also be started manually. It imports a shared Go source analysis, looks for files past a size threshold, and turns its findings into a single issue labelled refactoring and code-health. Issues expire after two days, and the workflow skips a run if a [file-diet] issue is still open, so it never piles up duplicates.

Its permissions are read-only for contents, issues and pull requests. The only write path is the issue it files through safe outputs.

What the logs show

The latest completed run, run 40338270065, finished successfully in about 9 minutes:

Outcome: success in one agent turn, with no reported workflow errors.
Output: one issue, #72111, flagging pkg/cli/token_usage_subagent_session.go at 1,557 lines against an 800-line threshold.
Network: the firewall recorded 20 requests, all allowed, none blocked.
Stability: against an earlier successful baseline (run 39956353556), turns and write-capable posture were unchanged.
Cost: the audit flagged medium “high token usage” at roughly 227k tokens, which is a reasonable thing to watch for a workflow that reads large files.

The issue it wrote

The report is more than “this file is big.” It maps the file into concerns: event schema types, unified token usage reading, session parsing, Copilot and Pi observers, error helpers, and row aggregation. It notes test coverage (a 645-line test file, roughly a 0.41 ratio) and proposes a split into focused files. That is the difference between a nag and a starting point: a maintainer or a coding agent can pick it up and begin.

Why it’s worth a look

Size thresholds are blunt, but a tool that applies them daily, explains the seams, and cleans up after itself keeps technical debt visible without creating noise. You can reproduce this evidence with gh aw logs and gh aw audit.

Try it

Browse the workflow sources, borrow the pattern for your own repository, and tell us what your agents find at github/gh-aw.

Weekly Update – October 26, 2026

Mon, 26 Oct 2026 00:00:00 GMT

It was a hardening week in github/gh-aw: six releases from v0.93.0 through v0.93.5 brought stronger static analysis, safer credential handling, and clearer failure diagnostics.

Release Highlights

The latest, v0.93.5, landed on October 25th. Earlier releases this week were v0.93.4, v0.93.3, v0.93.2, v0.93.1 and v0.93.0.

Network credentials in safe-output integration jobs (#71833): safe-output integration jobs can now use network credentials when they need them.
Engine driver crashes are classified (#71883): failures now come with secret-safe stderr diagnostics, so triage is faster.
Pi CLI 1.4.0 is the new default (#71899).
Sharper custom linters: a new readerstringconv analyzer (#71891), detection of parenthesized Flush discards (#71896), and missing recovery in goroutines launched through local closures (#71897).
Fewer surprises at compile time: indexing boundaries are guarded and cron-step overflow is prevented (#71889), and generated conclusion references and fetch warnings are fixed (#71898).

Notable Pull Requests

Prevent skipped threat detection from reporting inherited agent tokens keeps token reporting honest when detection is skipped.
Return stage-aware partial results for MCP compile and audit timeouts means a timeout no longer throws away the work already done.
Stop treating gh-aw-detection as experimental graduates threat detection from its experimental label.
Preflight agent assignment credentials before processing issues fails early instead of halfway through.

Agent of the Week: Daily Documentation Healer

The self-healing companion to the Daily Documentation Updater: it hunts for documentation gaps the updater missed and proposes fixes.

In its last three runs it ran at roughly 23:40 UTC each night, taking 7–8 minutes per run. The most recent run on October 25th succeeded, while the two before it failed in the conclusion job’s “Handle agent failure” step. The healer is also running a model-size experiment, alternating between Claude Sonnet 5 and Claude Haiku 4.5 to see whether the cheaper model heals docs just as well.

The healer is a documentation doctor that is itself part of a clinical trial, with a pre-registered hypothesis and everything.

Usage tip: Pair a generator workflow with a second “healer” pass that audits its output, so gaps get caught the next day.

→ View the workflow on GitHub

Try It Out

Update to v0.93.5 and tell us what you think. Feedback and contributions are always welcome in github/gh-aw.

Agent of the Day – October 23, 2026

Fri, 23 Oct 2026 00:00:00 GMT

Some code reviews catch bugs. Others catch cleverness, the extra abstraction or the helper used exactly once. Today’s spotlight goes to the workflow that specializes in the second kind.

Agent of the Day: Ponytail Reviewer

Ponytail Reviewer fires when a pull request is marked ready for review, skipping changes that only touch Markdown, docs or changesets. It can also be summoned on demand with a /ponytail or /review slash command in a PR comment. It runs on Codex, reads through a CLI proxy, and loads the ponytail-review skill, which carries the “does this need to exist?” rubric.

It never writes with broad permissions: contents: read and pull-requests: read are all it gets, and the review itself flows through the shared PR review base.

What the logs show

The most recent completed run, run 39965349799, reviewed a pull request on the pelikhan-silver-meme branch:

Outcome: success in about 6 minutes, one agent turn, no errors or warnings.
Footprint: roughly 108k tokens across five model invocations, with a peak input of about 27k tokens. The audit flagged this as “high token usage”, which is fair for a full diff read, and it makes a good optimization target.
Posture: read-only. The audit observed no GitHub write actions from the agent itself.
Network: the firewall saw 11 requests and blocked none.
Stability: compared with an earlier successful run (run 39880437110), the audit labelled it “stable”: same turns, same posture, same blocked-request count.

The audit also raised anomaly flags on a few log events. Those were rare-template matches, meaning patterns not seen before in this workflow’s history, rather than failures.

Why it’s worth a look

Reviewers are best at judgment calls, and a dedicated agent that only asks about complexity leaves them free for the rest. A narrow mandate, a tight permission set and a repeatable run profile make it easy to trust and easy to audit. You can pull the same data yourself with gh aw logs and gh aw audit.

Try it

Explore the workflow sources, copy the pattern and build your own focused reviewer at github/gh-aw.

Agent of the Day – October 22, 2026

Thu, 22 Oct 2026 00:00:00 GMT

Codebases rarely collapse in a single commit. They sag: a file grows a few hundred lines, a function picks up one more branch, an export list quietly doubles. Today’s spotlight goes to the workflow that notices the sag early.

Agent of the Day: Architecture Guardian

Architecture Guardian wakes up around 14:00 UTC on weekdays. It reads the commits from the previous 24 hours and looks for structural violations in Go and JavaScript files: large files, oversized functions, high export counts and import cycles. If it finds something, it files an issue prefixed [architecture-guardian] that expires after two days. If an earlier report is still open, a skip-if-issue-open guard stops it from piling on.

The tool surface is deliberately small. Bash is limited to cat, and the workflow carries a daily AI-credit cap and a three-denial limit, so a confused run ends quickly instead of wandering.

What the logs show

Recent runs, pulled with the logs and audit tools:

Run	Started (UTC)	Outcome	Tokens
#142	Oct 20, 14:21	Success, 7.9 min	6,941
#143	Oct 21, 14:23	Success, 10.9 min	7,030
#144	Oct 22, 14:22	Queued at time of writing	6,927 so far

Two things stand out. First, the footprint is remarkably steady: about 7,000 tokens per run, with four model invocations and almost no context rebuilding (a rebuild factor of roughly 1.001). Second, the audit of run #143 describes a linear path: one turn, two tool types, and zero created items. Its recorded output was a noop explaining that no Go or JavaScript source files had changed in the last 24 hours, so the scan was skipped.

That is the quiet virtue of this agent. A good guardian doesn’t need to raise an alarm every day. It checks, finds nothing worth reporting, says so explicitly and costs very little. The recorded output is auditable, so you can tell “nothing to report” apart from “silently broke”.

Note that the runs above did not create new issues or pull requests, so there are none to link this time. The evidence is the run history itself.

Takeaways

Pair scheduled agents with an explicit noop path so quiet days are visible, not ambiguous.
Narrow tools plus a credit cap make a cheap, predictable daily check.
Auto-expiring reports keep the issue tracker from filling with stale findings.

Want a guardian of your own? Explore the workflows, and compile one for your repository, at github.com/github/gh-aw.

Agent of the Day – October 21, 2026

Wed, 21 Oct 2026 00:00:00 GMT

Coverage tells you which lines ran. It says nothing about whether the test would ever notice a bug. Today’s spotlight goes to the workflow that asks the harder question.

Agent of the Day: Test Quality Sentinel

Test Quality Sentinel fires when a pull request becomes ready for review, or when someone types /review. A pre-fetch step collects the test diff, the new and modified Go and JavaScript test functions and some stats. The agent then reads that material with a tight read-only toolset (git diff, grep, cat, echo) and looks for tests that pin implementation details, cover only the happy path, inflate line counts or duplicate each other.

It also picks its model size per run. The audits below show experiment labels of claude-haiku-4.5 and claude-sonnet-5 side by side, so the cost of a review scales with the PR.

What the logs show

In the last few hours, the workflow ran five times on pull requests:

Run	Result	Duration
39613514534	Success	8.6 min
39614264303	Success	9.3 min
39614442224	Failure	5.1 min
39615043816	Failure	7 sec
39615569318	Still in flight in the log snapshot	12.3 min

I audited two of them. Run 39614264303 (on branch pelikhan-verbose-couscous) succeeded in two turns with no errors or warnings, a read-only posture and zero blocked requests. The audit compared it with the earlier run 39613514534 and flagged only the turn count (1 to 2) as a change. It also gave a medium-severity note for roughly 224k tokens in total.

Run 39614442224, on a purelock/lock-policy-functions-… branch, is the interesting one. It ended in failure after 5.1 minutes with two errors, one of them Agent reported that it could not complete the task. The audit classified it as risky and told me to investigate before treating it like the successful baseline. It still stayed read-only with nothing blocked. When the agent can’t finish, the workflow says so instead of posting a half-baked verdict.

Why it matters

A reviewer who only sees “92% coverage” can’t tell a sturdy suite from a decorative one. Sentinel gives a second opinion on test design at the moment a human is about to review. It is cheap, read-only and honest about failing. The 7-second failure in the table is a reminder that early-stage problems, such as pre-fetch or setup, are worth watching too.

Try it

Browse the workflow source, borrow the pre-fetch pattern for your own review agents, and star the repo: github/gh-aw.

Agent of the Day – October 20, 2026

Tue, 20 Oct 2026 00:00:00 GMT

Some workflows write code. Others decide what code gets written. Today’s Agent of the Day is Issue Monster , the Cookie Monster of the backlog: it nibbles on open issues and hands them to the Copilot coding agent, politely and in order.

Agent of the Day: Issue Monster

The Issue Monster runs every 30 minutes, and on demand via workflow_dispatch. Its frontmatter is mostly a set of manners:

Don’t pile on. If five or more open Copilot draft PRs (excluding those labeled broccoli) already exist, it skips the run.
Don’t start without food. skip-if-no-match requires at least one open issue.
Don’t build on a broken floor. skip-if-check-failing pauses it when the build, test, lint-go or lint-js checks are red.

A pre-step script searches for candidate issues and checks recent Copilot PRs for rate-limit signals, so the agent only sees a short, prepared list.

What the live run shows

I audited run 39438401745, a scheduled run on main from this morning:

Measure	Value
Conclusion	Success
Duration	7.9 minutes
Turns / tokens	4 turns, about 62k tokens
Errors / warnings	0 / 0
Firewall	19 requests allowed, 0 blocked
Engine	Pi with `copilot/gpt-5`

The agent made six safe-output calls: three assign_to_agent and three add_comment. The comments landed on #70227, #69063 and #68043. The single MCP server, safeoutputs, reported 6 requests and no errors.

The audit compared the run with an earlier successful run, 39433043735. Event, task domain and tool breadth all matched. Posture stayed write-capable and nothing was blocked. The one difference was the turn count, which the audit flagged as a change worth a glance rather than a problem. It also noted the token usage as medium severity, which makes sense given a roughly 36 KB prompt that carries the candidate list.

Why it matters

Handing work to a coding agent is easy; handing it the right amount of work is the craft. Issue Monster’s gates keep the queue of draft PRs short, avoid rate-limited moments and stay quiet while CI is failing. Its writes are narrow: comments and assignments, delivered through safe outputs.

Try it

Read the source, borrow the gating pattern, and let your own backlog get nibbled: github/gh-aw.

Agent of the Day – October 19, 2026

Mon, 19 Oct 2026 00:00:00 GMT

Every kitchen needs someone who walks the line, tastes things, and notices what’s gone cold. Today’s Agent of the Day is PR Sous Chef ‍, the workflow that does that for open pull requests.

Agent of the Day: PR Sous Chef

The PR Sous Chef is described in its own frontmatter as a workflow that nudges PRs idle for ten minutes with unanswered reviews and a branch update, without duplicate agent work. It fires every five minutes, and also on workflow_dispatch or a /souschef slash command on a pull request comment.

The trick is the gate. A skip-if-no-match search (is:pr is:open -is:draft, excluding Dependabot, Renovate and anything labeled broccoli) means the agent doesn’t spin up unless there is a PR worth looking at. Permissions are read-only, with a concurrency group so runs queue instead of overlapping. It runs the pi engine with copilot/claude-haiku-4.5, a small model for a small, frequent job.

What the live run shows

I audited run 39268406664, a scheduled run on main from this morning:

Measure	Value
Conclusion	Success
Duration	1.9 minutes
Errors / warnings	0 / 0
Posture	Read-only, no GitHub writes emitted
Behavior	Directed, narrow tool breadth, lean resources

The audit compared it with the previous successful run, 39264765467, from about 25 minutes earlier. It matched on event, task domain, execution style, resource profile and tool breadth, with no change in turns, posture or blocked requests. The audit classified the run as stable and recommended no action.

That is the point of this agent. A run that finds nothing to do, costs almost nothing and changes nothing is a healthy one. The same window also showed Architecture Guardian, Dead Code Removal Agent and Issue Monster in flight, so the Sous Chef is one of several automated helpers working in parallel.

Why it matters

Pull requests stall for dull reasons: an unanswered review comment, a branch that fell behind. Catching these within minutes, with a cheap model and a tight gate, keeps review loops moving without humans polling a queue.

Try it

Browse the workflow source, copy the gating pattern for your own repositories, and see what a quiet five-minute heartbeat can do: github/gh-aw.

Weekly Update – October 19, 2026

Mon, 19 Oct 2026 00:00:00 GMT

It was a busy week for github/gh-aw: five pre-releases, v0.92.1 through v0.92.5, landed between October 15 and 17. The theme was observability, security hardening and keeping engines current.

Release Highlights

v0.92.5

A security and diagnostics release.

Structured reports (#69756, #69755): generated issues now use a normalized report AST, and Actions-native and AW Markdown workflow ASTs were added.
Better failure visibility (#69788, #69467, #69432): sub-agent failures are surfaced, unavailable models are diagnosed, and audit/logs handle sub-agent instances without an agentId.
Safer MCP compile (#69510): MCP compile is read-only unless write tools are enabled.
Redaction (#69520, #69517): detection outputs and CLI proxy startup diagnostics are redacted before publication or artifact upload.
Dependency bumps to MCP Gateway v0.4.35 and firewall v0.28.62.

v0.92.4 and v0.92.3

You can now set create-issue: false to disable automatic issue output (#69369).
CLI-generated PR descriptions are richer, and there’s a new --pr-message flag (#69378).
gh aw upgrade now respects workflow engines and opt-out flags (#69362).
Copilot SDK sessions register MCP gateway servers, so MCP tools are available as expected (#69231).

v0.92.2 and v0.92.1

Every gh aw command that creates pull requests supports --auto-merge.
Engine upgrades: Copilot CLI 1.0.98, Codex 0.163.0, Gemini CLI 0.64.0, Goose 1.54.0, Kiro CLI 2.31.1 and Antigravity CLI 1.3.6 (#69146).
logs gained automatic failure clustering and investigation plans (#68954).
New offline skill command and SAML-safe instruction downloads (#69043).
health supports logs-style date windows (#69048).

Agent of the Week: daily-firewall-report

The network watchdog that collects firewall logs and reports on which domains your agentic workflows are reaching.

Its three most recent runs (October 6–8) were scheduled early-morning runs on the Copilot engine. Each ran for 10 to 27 minutes, and the firewall data it gathered showed 81 requests across the three runs, all allowed and none blocked, spread across api.githubcopilot.com, Sentry and Grafana ingest endpoints.

The funny part: a firewall reporter whose own firewall had nothing to block, while the runs themselves all stopped at the MCP Gateway start step. Even the watchers need a good night’s sleep.

Usage tip: Schedule a firewall report so unexpected domains show up in a daily summary instead of during an incident.

→ View the workflow on GitHub

Try It Out

Upgrade to v0.92.5 and tell us what you think. Feedback and contributions are always welcome in github/gh-aw.

Agent of the Day – October 16, 2026

Fri, 16 Oct 2026 00:00:00 GMT

Docs rot quietly. A feature ships on Tuesday, and by Friday nobody remembers that the reference page never heard about it. Today’s Agent of the Day is the workflow that keeps that from happening, and it also had a bumpy week.

Agent of the Day: Daily Documentation Updater

The Daily Documentation Updater wakes up around 10:00 UTC, reviews what merged recently, and opens a pull request when the documentation needs to catch up. It runs on the Codex engine and uses the create-pull-request safe output, so any change reaches the repo as a reviewable PR.

Here is what the logs show for the last three days:

Run	Started (UTC)	Duration	Result
38551459107	Oct 14, 10:42	10.2m	Success
38722208333	Oct 15, 10:42	37s	Failure
38888962279	Oct 16, 10:42	31s	Failure

The good run

The Oct 14 run used about 1.08 million tokens across 19 model invocations and about 197 AI credits. Its session shows it going through a batch of freshly merged work, including:

#68421: Show work titles and descriptions in work-queue issues
#68424: Add Anthropic Claude Haiku 5.5 pricing
#68426: Support numeric targets and improve diagnostic errors in push_to_pull_request_branch
#68428: Include AWF steering events in unified session audit reports

It also looked at the CLI consistency report, #68442, which is the kind of context that tells it where the docs and the CLI have drifted apart. That is the useful part of this agent: it connects “what merged” to “what the docs now owe readers”.

The rough mornings

On Oct 15 and Oct 16 the run ended in about half a minute. In both cases the audit data points to the same place: the activation job failed at the step Restore daily AIC scan observations, before the agent started. The agent never ran, so these are infrastructure failures rather than a documentation mistake. They are worth fixing because a daily workflow that silently skips two days is a doc backlog in the making.

Cost at a glance

The successful run’s friction data is a useful reminder that long sessions rebuild context: peak input was about 79k tokens, but cumulative input reached roughly 1.07M. Trimming what the agent re-reads is the obvious lever for the next optimization.

Try it yourself

Want docs that follow your merges? Browse the workflow source, copy the pattern, and adapt it to your repo at github/gh-aw.

Agent of the Day – October 15, 2026

Thu, 15 Oct 2026 00:00:00 GMT

Code explains what changed. Almost nothing explains why. Today’s Agent of the Day is Design Decision Gate , the workflow that stands at the pull request door and asks for the reasoning.

Agent of the Day: Design Decision Gate

Design Decision Gate runs on pull requests. It enforces Architecture Decision Records (ADRs) before implementation work merges: it inspects the PR, decides whether a significant design decision is hiding in the diff, and, using the ADR-writer agent, drafts one when it’s missing.

Here is what the live logs show for the last few hours, all three runs successful:

Run	Started (UTC)	Duration	Tokens
38743544914	Oct 15, 13:32	8.8m	4,977
38745396329	Oct 15, 13:45	7.3m	432
38745652557	Oct 15, 13:47	9.0m	4,223

The third run is the interesting one. On PR #69048, the gate pushed a commit (cda3028) to the PR branch and then left a comment explaining what it did. Not a nag, but an actual contribution the author can edit or reject.

The middle run is a nice contrast: just 432 tokens. When a PR has nothing architecture-shaped in it, the gate has little to say and spends almost nothing saying it.

Why it matters

Decisions get written down while context is fresh, not six months later during an archaeology session.
Authors get a head start, since a draft ADR is far easier to correct than a blank page.
Cost scales with relevance: trivial PRs pass cheaply, design-heavy ones get real analysis.

Other agents were busy on the same PR, too. Code-quality and skills reviewers were commenting on #69048 within minutes of the gate, a fair picture of how several small specialists can cooperate on one change.

Try it

Browse the workflow source, adapt the ADR rules to your own team, and let your pull requests explain themselves: github/gh-aw.

Agent of the Day – October 14, 2026

Wed, 14 Oct 2026 00:00:00 GMT

Some of the best teammates are the ones who only speak up when it matters. Today’s Agent of the Day is Architecture Guardian, a scheduled workflow that watches code structure so humans don’t have to.

Agent of the Day: Architecture Guardian

Every weekday around 14:00 UTC, Architecture Guardian wakes up, reads the commits from the previous 24 hours, and checks Go and JavaScript files for structural trouble: oversized files, bloated functions, packages exporting too much, and import cycles. It runs on the Copilot engine with read-only permissions and a daily credit cap. A skip-if-issue-open guard stops it from piling duplicate reports on top of an unresolved one.

What do the live logs say? Over the last week the workflow completed five consecutive runs, all successful:

Run	Date (UTC)	Duration
38579847583	Oct 14	7.3m
38413487304	Oct 13	7.8m
38263495310	Oct 12	6.5m
37944177840	Oct 9	7.3m
37792402560	Oct 8	9.6m

The aggregate numbers are pleasantly dull: no errors, no warnings, no missing tools, no agent-logic failures, and zero safe-output items across the window. The four completed runs with token data each used about 6k tokens in a single turn, for roughly 7–9 AIC apiece. In plain terms, the codebase stayed within the guardrails, so the agent had nothing to file.

That silence is the feature. A noisy linter trains people to ignore it. A watchdog that stays quiet on healthy days and opens a [architecture-guardian] issue only when a real violation appears, with a suggested fix attached, earns attention when it does speak.

One caveat: the safe-output count is zero across these five runs, so this window shows the “all clear” path rather than a filed finding. Earlier architecture-guardian issues exist in the tracker if you want to see what a flagged violation looks like.

Try it yourself

Want a structural watchdog in your own repo? Browse the source, copy the pattern, and compile your own with gh aw. Find the project at github/gh-aw.

Weekly Update – October 12, 2026

Mon, 12 Oct 2026 00:00:00 GMT

Five releases shipped in github/gh-aw this week, from v0.91.4 to v0.91.8. The themes: smarter model routing, new engine options, and sturdier sandboxes.

Release Highlights

v0.91.8

Agent sessions now render as full elapsed-time timelines, with consistent trace normalization across Copilot, Codex, Claude, Gemini, Pi, Kiro, Aider and AGY.

Built-in graders: a new code-read-precision trajectory grader and builtin grader defaults loaded from setup-action JSON.
Model routing: Copilot wire APIs are selected from AWF metadata and model mismatches are surfaced. The firewall moved to v0.28.51.
Hardening: a custom grader sandbox escape was closed, and workflow YAML, scanner file handling and temp-file writes were tightened.
Memory validation: declarative JSON schema validation for memory, with repairable invalid repo-memory baselines.

v0.91.7

This release introduces a native Go Agy engine (#66998) and deprecates the Gemini engine in its favor (#66941), so plan to migrate Gemini workflows.

Work queue upgrades: priority-aware conformance dispatch (#67222) and scoped GitHub Issues mirrors (#67025).
Better model visibility: routed model and effort are reported across outputs (#66963), plus per-agent spend (#66961).
Secret detection now covers all AWF environment sources (#67017).

v0.91.6

Git-backed fair DAG work queues with claim-scoped effects (#66024).
Model routing across Claude, Codex, and pi (#66723).
Stricter pull_request_target checkouts via an explicit strict-mode policy (#66649).

v0.91.5

Ordered engine model fallback keeps workflows running when a preferred model is unavailable (#66593).
auto and copilot/auto models now work in Pi, Codex, and Claude (#66567).
New gh aw fix codemods for Copilot web-fetch and reserved inputs (#66556).

v0.91.4

v0.91.4 adds support for Claude Code with GitHub Copilot inference and local-first workflow debugging.

Agent of the Week: unbloat-docs

Documentation Unbloat reviews the docs and simplifies verbose pages while keeping them clear and complete.

This week it ran every night on its schedule at about 01:15 UTC. All three runs we looked at (October 10, 11 and 12) succeeded, taking between 7 and 10 minutes and producing one safe output each.

Its stated mission is to make docs shorter, so it spends its nights trimming words, which is a bit like a barber who works only while you sleep.

Usage tip: Pair a scheduled editor like this with a /unbloat slash command so you can also trim a specific page on demand.

→ View the workflow on GitHub

Try It Out

Upgrade to v0.91.8 and see the full release list. Feedback and contributions are welcome in github/gh-aw.

Weekly Update – October 5, 2026

Mon, 05 Oct 2026 00:00:00 GMT

Three releases shipped in github/gh-aw this week, headlined by work queues and built-in ledgers in v0.90.3. Here’s what’s new.

Release Highlights

v0.90.3

This release introduces work queues and ledgers, tightens safe-output guardrails, and broadens Copilot model support.

Work queue & ledgers: built-in log, set, map, table, and counter ledgers, plus Git-backed work-queue operator commands. See the ledger compaction docs. Related PRs: #65494 adds selectable issue-backed queue storage, and #65443 dispatches queued work to trusted worker workflows.
Copilot engine: AWF task-level model routing, none and max reasoning efforts, GPT-6.1 Sol support, Claude 5.5 and GPT-6 pricing, and a run-wide tool-call budget for Copilot SDK agents.
Safe outputs: add-labels supports separate call and per-call label limits, and several hardening fixes landed. #65612 adds attributed issue creation for custom safe-output jobs.
Docs: a redesigned site with a new landing page, warm palette, and Mona Sans.

Heads-up on behavior changes: GitHub App tokens now require explicit scopes at compile time, Claude defaults to strict MCP configuration, and add-comment fails when the triggering target can’t be resolved.

v0.90.1

Standalone safe-output-backed ledgers arrived (#64354), along with richer audit artifacts covering ledger transactions and threat-detection outcomes (#64509, #64506), and compile-time self-hosted runner enforcement (#64173).

v0.90.0

Step summaries now render friction cost, agents may emit two noop calls by default (#64031), and five new trajectory graders were added.

Other Notable Pull Requests

#65423 adds a TLA+ workflow security model and fail-closed credential cleanup.
#65564 emits incomplete outcomes with diagnostics when an agent produces no safe outputs.
#65650 makes PR Sous Chef poll faster and start agents only for actionable work.

Agent of the Week: daily-arxiv-researcher

A daily scout that reads the latest arXiv papers and looks for ideas that could improve GitHub Agentic Workflows.

Its recent runs have all succeeded in about five to six minutes. The October 2 and 3 runs used the Copilot CLI engine, and the October 4 run was the first on the Codex engine with web search and fetch tools, after a recent change moved it there. That run used roughly 84K tokens, versus 7K and 13K for the earlier runs.

The new engine is evidently more curious: it read about six times as many tokens as its predecessor on the same morning schedule.

Usage tip: Pair a research workflow with a dedup cache, as this one does with cache-memory, so it never reports the same paper twice.

→ View the workflow on GitHub

Try It Out

Update to v0.90.3 and give work queues a spin. Feedback and contributions are welcome in github/gh-aw.

Agent of the Day – September 28, 2026

Mon, 28 Sep 2026 00:00:00 GMT

Every repository accumulates issue clutter — related bugs filed separately, follow-ups scattered across weeks, nobody quite sure which ticket is the “real” one. Today’s Agent of the Day is the workflow built to fight that entropy: Issue Arborist, a daily codex-powered agent that reads the last 100 open issues in github/gh-aw and cultivates order out of them.

Agent of the Day: Issue Arborist

Issue Arborist’s job description is disarmingly simple: fetch open issues without a parent, spot the ones that belong together, and link them as sub-issues under a fresh tracking parent. No opinions about priority, no code changes — just topology. It runs once a day on a schedule, reads with github: mode: local in read-only issue scope, and writes exclusively through safe outputs (create-issue, link-sub-issue, create-discussion), so every action it takes is logged, reviewable, and reversible.

Its most recent run on September 28 is a good showcase of both the value and the honesty built into these agents. Working from 63.2k tokens and 11.6 AIC of compute, Issue Arborist scanned the open backlog and found two clusters worth grouping:

#63931 – “AW workflow failure and incomplete-result incident tracking”, a parent stitching together 33 recurring incident issues from daily workflow runs — everything from failed follow-ups to no-safe-output cases — into one place a maintainer can actually triage instead of scrolling past one-off tickets.
#63932 – “Deep-report quick-win remediation backlog tracking”, grouping 12 smaller quick-win fixes flagged by the deep-report workflow into a single actionable backlog.

It also opened discussion #63933 in the audits category to summarize the pass, following the same reporting convention every gh-aw daily agent uses so a human skimming discussions gets the same story without opening 45 issues.

Here’s the part that makes this a genuinely interesting pick rather than a tidy success story: the run’s audit trail shows it attempted 34 link_sub_issue calls and most of them failed with Target is "triggering" but not running in issue context, skipping link_sub_issue — a context mismatch between the scheduled trigger and the sub-issue linking tool. The workflow’s own audit report flags this plainly as a workflow_failed critical finding, not something smoothed over. And yet the two parent issues and the discussion still landed cleanly, because gh-aw’s safe-outputs model treats each output independently — a broken tool call doesn’t roll back the ones that already succeeded. Compare that to the previous day’s run, which completed with no findings and no issues to group, a perfectly quiet, uneventful pass.

That contrast is the whole point of running the same agent daily: some days there’s nothing to say, and some days there’s a real cluster to surface — and when a downstream tool call breaks, gh-aw’s audit tooling makes sure the failure is visible instead of buried in a green checkmark. Issue Arborist doesn’t pretend the run was flawless. It reports the win and the wart in the same breath, which is exactly what you want from automation touching your issue tracker unsupervised.

Want to see the compiled workflow, the safe-outputs config, or borrow the pattern for your own repo? Start here: github/gh-aw.

Weekly Update – September 28, 2026

Mon, 28 Sep 2026 00:00:00 GMT

It was a big week for github/gh-aw: three releases shipped, a legacy sandbox runtime was retired, and the firewall got noticeably stricter about what agents can reach. Here’s the rundown.

Release: v0.89.22

v0.89.22 landed on September 27th as the week’s headline release, building on v0.89.21 and v0.89.20 from earlier in the week.

! Breaking Change

Docker sbx and gVisor sandbox runtimes removed (#63034) — isolated agent execution now runs exclusively on Cloud Hypervisor. If your workflow frontmatter still sets sandbox.agent.runtime: docker-sbx or gvisor, switch it to cloud-hypervisor before you next recompile.

What’s New

Hosted-web domain policies in frontmatter (#63212) — the new network.hosted-web key lets you allow or block domains that provider-hosted Claude and Codex web tools reach outside AWF’s network boundary.
Firewall enforcement got sharper — Copilot and web tools now get enforced firewall compatibility checks (#63474), scoped precisely to the web tools a workflow actually enables (#63632), while false blocked-domain warnings for otherwise-successful requests were squashed (#63246).
Grouped audit findings mode (#63032) — gh aw audit --group now aggregates findings by run and code with occurrence counts, in pretty, markdown, or JSON output.
More audit visibility — an opt-out for automatic audit baseline downloads (#63012) and MCP payload size reporting in audit (#63687) give workflow authors finer control over what audit runs measure and fetch.
Large Copilot prompts streamed through stdin (#62766, thanks @davidslater!) — prompts over 100 KiB are now delivered via stdin instead of being truncated, preserving full context for big agentic workflows.
Repo-memory backend for daily AIC guardrail (#62958) — memory persistence now supports repository-backed storage for cost accounting.

Notable Fixes

Fixed dangling safe-outputs-app-token references when safe outputs are staged (#63490).
Fixed repo memory retry head refresh authentication (#63497) and prevented cache-memory validation marker EACCES failures (#63498).
Pinned pull_request activation checkout to the base SHA for improved supply-chain safety (#63499).
Fixed slash-command PR prefetch caching and error handling (#63678).

Notable Pull Requests

Add friction-cost attribution to audit and usage artifacts — audit output now attributes friction costs directly, making it easier to spot where workflows are burning extra cycles.
Track skill usage in audit artifacts — you can now see which skills workflows actually invoke, straight from the audit trail.
Surface AWF steering counters in compact audit usage data — steering activity is now visible in the compact usage summary, not just the full report.

Agent of the Week: ci-coach

Meet ci-coach, the CI Optimization Coach that runs daily to hunt down slow, imbalanced, or wasteful spots in gh-aw’s CI pipeline and proposes fixes as pull requests.

This week ci-coach had its hands full with cgo.yml’s notoriously lopsided unit-test shard. On September 24th it noticed the alphabetic A-C shard was taking ~106 seconds — more than double the other four shards’ 40-48 seconds — and traced nearly 20% of that gap to a single test full of pointless time.Sleep-backed cooldowns. It opened a fix (#63187) that reconfigured the mocked rate limit so the test skips its legacy 500ms sleep loop entirely. The very next day, still not satisfied, it went after the same shard again — this time proposing a full matrix rebalance that carves out a dedicated “Linters” shard, projected to cut the A-C shard’s time from ~93s down to ~29s.

Two days, two PRs, one stubbornly slow test shard — ci-coach is basically the CI equivalent of someone who keeps rearranging the furniture until the room finally feels right.

Usage tip: Run this class of workflow on a schedule against your CI config files — it’s great at spotting shard imbalances and dead-weight sleeps that are easy to miss by eye but add up fast across hundreds of runs.

→ View the workflow on GitHub

Try It Out

Update to v0.89.22 today, double-check your sandbox runtime settings if you were on docker-sbx or gvisor, and give the new gh aw audit --group mode a spin. As always, questions and contributions are welcome in github/gh-aw.