AAM Labs — Execution Preview

Agents, pipeline and real runs · data as of 24 Oct 2026

1. Run an idea for real

REAL RUN — real model, real sandbox, real tests

Runs on your assistant's computer (not a paid always-on server). If that computer is off or restarting, runs will wait. Cost cap $0.05 per job and $0.50 per day, one job at a time. A pull request is opened only when you tap Approve.

Access code (kept only in this browser)

1b. Try an idea (simulated)

This is a static preview. Pressing “Preview run” plays a simulated walkthrough in your browser only. Nothing is sent or stored, and no agent runs. For a real run use the section above.

2. The agents

Five LangGraph roles in the orchestrator:

supervisortakes the requirement, pauses at the owner gate
architecturewrites the plan
autonomy_controllercreates the branch and sandbox, enforces the budget and iteration limits
developmentedits code inside the sandbox (gpt-4o-mini via LiteLLM)
testingruns pytest and the security scans

State machine:

intake → plan → branch → implement → test → (fail: back to implement, max 3) → security → owner_gate → PR

Sandbox: Docker, no network, non-root, all capabilities dropped, read-only filesystem. Budget cap $0.05 per job.

3. Real work done

Repo: 37ma1178-chin/aam-experiments (private). Model: openai/gpt-4o-mini. Each task passed on the first attempt.

PR #1 — Fix off-by-one in truncate
6 passed · gitleaks/trivy/bandit passed · $0.000207
base: mainOpen PR #1

PR #2 — Add slugify()
7 passed · gitleaks/trivy/bandit passed · $0.000368
base: PR #1 branchOpen PR #2

PR #3 — Input validation for mean() and clamp()
8 passed · gitleaks/trivy/bandit passed · $0.000201
base: PR #1 branchOpen PR #3

PR #4 — Docstrings and README usage section
6 passed · gitleaks/trivy/bandit passed · $0.000240
base: PR #1 branchOpen PR #4

Total model cost for the four tasks: about $0.001 ($0.001016).
Orchestrator test suite: 15 passed.
Repair loop demo (stand-in model, not a real model): iteration 1 failed 2 tests, iteration 2 passed 6/6.

4. Status and caveats

No PR is merged yet. Merge #1 first, because #2–#4 are based on it.
PR #1 behaviour change: truncate("ab", 2) now raises ValueError.
The repair loop has only been proven with the stand-in model. The real model got every task right on the first try.
All four jobs still show APPROVED_PR_PENDING in Postgres, not PR_CREATED.
The CLI's push-and-PR path hasn't run against GitHub yet. The PRs were opened by a separate step.
Real runs execute on the assistant's computer through a relay on this site. It is not a paid always-on VM.