North star: amux owns execution, state, isolation, recovery, observability, and verification. The model owns reasoning.
Positioning: amux is not another agent — it is the durable operating system around agents. Every item here makes any model operate in a cleaner, safer, more information-rich environment, and gets better automatically as models improve. We deliberately avoid elaborate planners, proprietary agent loops, and prompt-heavy coordination that rots as models get stronger.
Triage outcome (from a 20-proposal review)
Build — endorsed (own issues, linked below): agent runtime contract + capability registry · event-sourced session state · automatic worktree/env isolation · verification gates + checkpoints + retry · structured task & handoff envelopes · leases & resource reservations · exactly-once & transactional side effects · deterministic policy + trigger engine · dependency-aware task DAG · MCP broker (client + server) · OTel tracing + provenance + harness metric · harness evaluation suite · capability-based permissions.
Build but scoped: context assembly broker — ship the deterministic, inspectable core (provenance, pinning, change-invalidation, budget); defer ML relevance-ranking, the piece most likely to be obsoleted as models gain longer context + native retrieval.
Deferred (no issue yet): external A2A interoperability — expose an A2A boundary only once amux needs cross-platform agent exchange beyond its own sessions; the internal event/task protocol stays simpler and more controllable.
Explicitly NOT building (anti-patterns that rot as models improve)
A proprietary ReAct loop · role-play personas · fixed researcher→planner→coder→critic pipelines · LLM-generated coordination chatter · provider-specific prompt optimization · amux-controlled chain-of-thought · auto task-decomposition with no human-visible graph · a custom tool protocol (MCP suffices) · a custom agent-comms protocol (until a real cross-platform need).
Recommended sequence
- Runtime contract + capability registry → 2. Event-sourced session state → 3. Worktree/env isolation → 4. Verification gates + checkpoints + retry → 5. Task + handoff envelopes → 6. MCP broker → 7. DAG + leases + resource locks → 8. OTel traces + harness evals → 9. Transactional side effects across non-code workflows → 10. External A2A (later).
Product metric
Not tokens — verified tasks completed per unit of human attention, time, and cost.
Sub-issues are appended below once created.
Sub-issues
Foundation
Execution & verification
Coordination (deterministic)
Tools / interop
Observability
Safety
Scoped / revisit
For contributors (seams-first)
Seams — maintainer-owned, land first (everything else plugs into these): #47 (agent runtime contract + capability registry), #48 (event-sourced session state).
Leaves — contributors welcome (help wanted): #49–#60. Provider adapters, verification runners, MCP tools, eval scenarios, and policy hooks are the most parallel-friendly. Many leaves unblock once the seams exist — each issue's Depends on line says what it needs.
To pick one up: comment to claim it, keep PRs small and single-region, and rebase before pushing (it's one ~50k-line file). See CONTRIBUTING → Building the harness roadmap.
North star: amux owns execution, state, isolation, recovery, observability, and verification. The model owns reasoning.
Positioning: amux is not another agent — it is the durable operating system around agents. Every item here makes any model operate in a cleaner, safer, more information-rich environment, and gets better automatically as models improve. We deliberately avoid elaborate planners, proprietary agent loops, and prompt-heavy coordination that rots as models get stronger.
Triage outcome (from a 20-proposal review)
Build — endorsed (own issues, linked below): agent runtime contract + capability registry · event-sourced session state · automatic worktree/env isolation · verification gates + checkpoints + retry · structured task & handoff envelopes · leases & resource reservations · exactly-once & transactional side effects · deterministic policy + trigger engine · dependency-aware task DAG · MCP broker (client + server) · OTel tracing + provenance + harness metric · harness evaluation suite · capability-based permissions.
Build but scoped: context assembly broker — ship the deterministic, inspectable core (provenance, pinning, change-invalidation, budget); defer ML relevance-ranking, the piece most likely to be obsoleted as models gain longer context + native retrieval.
Deferred (no issue yet): external A2A interoperability — expose an A2A boundary only once amux needs cross-platform agent exchange beyond its own sessions; the internal event/task protocol stays simpler and more controllable.
Explicitly NOT building (anti-patterns that rot as models improve)
A proprietary ReAct loop · role-play personas · fixed researcher→planner→coder→critic pipelines · LLM-generated coordination chatter · provider-specific prompt optimization · amux-controlled chain-of-thought · auto task-decomposition with no human-visible graph · a custom tool protocol (MCP suffices) · a custom agent-comms protocol (until a real cross-platform need).
Recommended sequence
Product metric
Not tokens — verified tasks completed per unit of human attention, time, and cost.
Sub-issues are appended below once created.
Sub-issues
Foundation
Execution & verification
Coordination (deterministic)
Tools / interop
Observability
Safety
Scoped / revisit
For contributors (seams-first)
Seams — maintainer-owned, land first (everything else plugs into these): #47 (agent runtime contract + capability registry), #48 (event-sourced session state).
Leaves — contributors welcome (
help wanted): #49–#60. Provider adapters, verification runners, MCP tools, eval scenarios, and policy hooks are the most parallel-friendly. Many leaves unblock once the seams exist — each issue's Depends on line says what it needs.To pick one up: comment to claim it, keep PRs small and single-region, and rebase before pushing (it's one ~50k-line file). See CONTRIBUTING → Building the harness roadmap.