Autonomous task agent
An agent that reads a backlog, plans it, executes it under a spend cap, and is graded by something other than itself.
-
Nobody hands it the work.
It reads any tasks.md under the workspace — opt in or opt out — and plans what it finds. There is no queue to fill in and no one to assign anything.
-
It spends real money, under a cap.
Planning and execution run against a daily cost cap and an approval queue, and it reports cost and tokens per day, week and month. The cap is enforced, not requested.
-
It does not mark its own homework.
A task is finished when the project's own tests say so, not when the model says so. The verifier's result goes to an append-only ledger, and repeated failures stop the loop.
- 01 Designed an autonomous agent that discovers project backlogs (any tasks.md under a workspace, opt-in/exclude), plans and executes tasks under a daily cost cap and approval queue, and reports cost/tokens per day, week, and month via a status dashboard
- 02 Built a verifier-gated task loop that grounds task completion in real test runs instead of LLM self-report, backed by an append-only feedback ledger and a consecutive-verify-failure cap.
- 03 Hardened an autonomous coding agent's execute mode with a Bash command denylist and a base-branch-move guard, closing the last safety gap before unattended runs.
- 04 Diagnosed and repaired five silent merge-conflict regressions that had broken a Python package's imports and disabled 45 tests at collection, by comparing each merge commit against its parents to restore the correct code.
Interface recreated with invented data — no production screenshot is published. Private system; walkthrough and code review available on request.
- Role
- Sole engineer
- Timeline
- 2026 — ongoing
- Stack
- Python · LLM orchestration · pytest