Run Claude Code and Codex as a real engineering system—independent review, isolated worktrees, verification before approval, and a complete evidence trail after every run.
npm install -g @devin-lai/aios
Why AIOS
Most agent workflows optimize for producing code. AIOS optimizes for producing code with enough evidence to trust.
The author reviews its own work.
A different engine must review every change.
Parallel edits share mutable state.
Every task gets a dedicated branch and worktree.
Tests are an optional follow-up.
Verification failures are blocking review issues.
Failures disappear into chat history.
Prompts, responses, checks, and verdicts stay on disk.
The system
AIOS turns agent work into a bounded, inspectable delivery pipeline. Every transition has a reason, an artifact, and a failure state.
Claude and Codex draft independently, then merge, polish, and critique the plan.
output / project.mdThe specification becomes dependency-aware tasks with explicit acceptance criteria.
output / tasks/*.mdEach task runs on its own branch in its own worktree. Your checkout stays yours.
state / aios/task/<id>Tests, lint, type checks, and builds become evidence for an independent reviewer.
gate / all greenAfter rebase, AIOS runs review and verification again against the actual merged tree.
gate / re-reviewThe staging branch becomes a pull request, waits for CI, and squash-merges on success.
result / mainEvidence, not vibes
A run is not a disappearing conversation. AIOS writes the full chain of evidence locally so failures can be diagnosed without rerunning the agents.
$ aios ship "Add request IDs with tests"
✓ spec synthesized 18.4s
✓ 4 tasks scheduled parallel: 3
✓ test · lint · build all green
✓ independent review approved
✓ post-rebase gates green
✓ merged PR #42 main.aios/runs/2026-07-27T07-22-18/
├── specgen/
│ ├── 1-draft-claude.md
│ ├── 2-draft-codex.md
│ └── 5-critique.md
└── 002-request-middleware/
├── round-1/
│ ├── coder.prompt.txt
│ ├── coder.response.raw
│ ├── verify.json
│ └── reviewer-response.json
└── report.md[NEEDS HUMAN] stall_no_progress
task: 003-migrate-schema
unmet: rollback coverage
blocking: NOT NULL column has no default
Work preserved on:
aios/task/003-migrate-schema
Resume: aios unblock 003-migrate-schemaBuilt to stop safely
AIOS spends its complexity on the places where unattended coding usually fails: isolation, cancellation, stale evidence, secrets, and loops that should have stopped.
Independent review
Configuration and runtime selection both reject identical coder and reviewer engines. The separation is a program invariant, not a prompt suggestion.
coder_default != reviewer_default
enforced
Isolation
Task work and merge integration happen in dedicated worktrees.
Bounded recovery
Retry, revise, decompose, respec, then structured human handoff.
Scoped tools
Tasks opt into named servers; retained configuration is redacted.
Cost visibility
Inspect per-run estimates and recurring review patterns locally.
Start in three commands
AIOS ships as a native binary through npm. You need Git 2.40+ and authenticated Claude Code and Codex CLIs. GitHub CLI is required for the pull-request flow.
npm i -g @anthropic-ai/claude-code @openai/codex
npm i -g @devin-lai/aios
aios init && aios doctor && aios
Open source under MIT
Keep the agents you already use. Add the separation, verification, and evidence that autonomous work needs.