Skip to content

Repository files navigation

Augustus

Find, build, evaluate, and improve systems with decision models.

License: MIT Release Quality Docs

Augustus equips agents to find useful placements, build decision-driven systems, create evaluations, and hill-climb them against real outcomes. Reach for it when a step could be code, an LLM call, a classifier or ranker, or a person; when a confidence score needs a threshold, abstention, or a human fallback; or when a prompt or program needs an eval before you optimize it. It draws on decision theory, value of information, multi-criteria analysis, signal detection, search/control, and formal-methods boundaries across software, business, organizations, research, and everyday decisions.

TypeSafe Jev (Choice, Score, Noul) is the default hosted exemplar. The skill also covers classical classifiers, encoders, open decision heads, constrained readouts, rankers, and vision scorers. Choose the family by the task, then test it against the baseline. Sometimes the best result is a formula, a checklist, or no new model.

The working model is:

evidence → bounded judgment → explicit policy → checked action → observed outcome

Augustus is independent of TypeSafe. The official TypeSafe skill and current provider docs own API contracts; Augustus supplies composition, implementation, evaluation, and improvement methods. Named for Augustus De Morgan, mentor of William Stanley Jevons.

Jev & Augustus, a 2:38 cut-paper film: where should your agent's judgment go?

Watch the 2:38 film: Jevons' logic piano, De Morgan's boundary, and how Augustus places and tests a bounded judgment.

Try it

After installation, ask your agent:

Use Augustus to audit our refund-email workflow. Find the smallest useful classifier insertion, keep eligibility and payments in code, and propose an evaluation that could reject the change.

Use Augustus to compare ways our library could choose three programs under a fixed budget. Make the values, evidence gaps, and tradeoffs explicit.

Use Augustus to review this confidence threshold. Explain what the score means, when to abstain, and what we should measure on held-out cases.

Use Augustus to build a decision-model router and its evaluation harness. Keep our incumbent runnable, test complete episode outcomes, and set up a bounded improvement loop with untouched confirmation data and rollback.

Advice produces a concise design card and falsifier. Build requests produce working adapters, policy and evals; improvement requests produce a bounded incumbent–challenger loop. The agent reads only relevant references. Source popularity and proxy-score gains do not establish improvement.

Examples

Problem Placement Evaluate
Expensive generated-JSON email routing Bounded intent classifier before existing handlers Action errors, review coverage, total cost
Search results need ordering Retrieve candidates, then rank relevance Recall, nDCG, final task success
Many plausible projects under a budget Explicit utility/MCDA with exact constraints Sensitivity, feasibility, stakeholder outcomes
Agent claims it is finished Judge evidence gaps; verify artifacts and effects False completion and recovery on real tasks
Need a decision under uncertainty Compare act, defer, and gather-more-evidence Expected loss and value of information
Model appears to approve a risky action Treat judgment as evidence inside host policy Unauthorized effects, failure paths, drift

A typed response is not proof of truth. Ranking scores, probability, confidence, calibration, and action success have different meanings. See the working skill.

Install

Install the published release, 0.7.2:

# Claude Code
claude plugin marketplace add 24601/Augustus@v0.7.2
claude plugin install augustus@augustus

# Codex, Cursor, and other Skills CLI agents
npx skills add https://github.com/24601/Augustus/tree/v0.7.2/.agents/skills/augustus

To follow the default branch, which may contain unreleased -dev work, use claude plugin marketplace add 24601/Augustus or npx skills add 24601/Augustus --skill augustus. A pinned install stays on its tag. To move a Claude Code install to another tag, run claude plugin marketplace remove augustus, then add and install again. For a manual install, copy .agents/skills/augustus/ with its references and scripts; Codex reads user skills from ~/.agents/skills/.

The skill needs no API key to provide design guidance. Calling Jev or another hosted provider is a separate, optional integration with its own credentials and costs. Review installed instructions before granting any agent access.

In Claude Code, invoke /augustus (a plugin install also answers to /augustus:augustus); in Codex, use $augustus. With many skills installed, a host may shorten or drop skill descriptions, so name the skill explicitly if it is never chosen. If it is not visible, reload your agent's skills/plugins. Check the installed version with claude plugin details augustus@augustus or metadata.version in the installed SKILL.md. See worked examples for the kind of result to expect. Avoid installing the same skill by multiple methods in one agent.

Project and evidence

  • Skill and reference index: runtime guidance.
  • Research archive: primary sources, historical claims, revisits, and the current decision-model review.
  • Contributing: content boundaries, tests, behavioral review.
  • Changelog: product changes, separate from research observations.
  • Website: examples and ecosystem orientation.
  • Feedback: installation problems, mistaken activation, and sanitized real-world failures.

For local development, install requirements-dev.txt and run make check. The offline helpers evaluate labeled binary predictions and paired workflow outcomes; neither calls a model. The composition calculus and build/improvement workflow connect methods to implementation and evidence. Tests and structural lint do not establish model quality or deployment benefit.

Versioning

Published release: 0.7.2. See the release notes for changes and migration details. The default branch is now 0.8.0-dev: unreleased work toward a second skill, augustus-train. Install from the v0.7.2 tag when you need the published revision.

Historical TypeSafe skill provenance: v0.5.7 (65a39f3), rechecked on 2026-09-23 as that repository's latest tag and HEAD. Read live provider docs before writing integration code; that pin is not a current API guarantee.

License

MIT. See LICENSE. Security reports: SECURITY.md. Contributions: CONTRIBUTING.md.

About

Agent skill for finding, building, evaluating, and improving decision-model systems. Compositional decision calculus, eval harnesses, and bounded prompt/program hill climbing for Software 3.0. TypeSafe Jev is the default hosted exemplar. Independent of TypeSafe.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

10 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages