Overcontrol Software Factory
A system for running autonomous coding-agent work as a supervised factory: a generic run driver parameterized by a frozen run request, deterministic lane supervision with interrupt and bounded recovery, independent evaluation gates that do not accept an implementer's own confidence as completion evidence, and durable agent policy carried in the repository rather than re-steered each session.
Problem
A coding agent that reports success is not evidence that the work is correct, and an agent that is stuck produces no signal at all. Orchestration assembled per session leaves no record of what policy a run executed under, and quality gates chosen by the agent's own judgment vary with how hard it happened to be thinking.
Solution
Built the run driver, supervision, and gate policy as versioned artifacts the run verifies before trusting them, and measured the workflow itself — model, effort, and gate configuration — as controlled experiments whose results changed the defaults.
Role
Architect and engineer
Technologies
Impact
- Replaced hand-derived per-run operators with one generic run driver parameterized by a frozen run request, validated by building the same ticket both ways: the composed launch instruction matched at 0.995 similarity, differing only in run id, timestamp, and profile name
- Ran controlled agent experiments on identical workloads and commits, measuring reasoning effort at 43.4 minutes and 298k output tokens against 15.5 minutes and 49k, then found the cheaper arm had silently skipped the security review — nine static-analysis findings against three — and made gates repository-configured rather than effort-dependent
- Built deterministic stall detection after an agent lane spent seventy minutes and 167k output tokens re-editing one file against an impossible hypothesis that no quality gate could see because it never produced a commit, using repeated identical tool failures, per-lane output-token caps, and time since the last reachable commit
- Shipped stall detection whose recovery path was unreachable on every run that spawned a subagent, found when it first fired live, and closed the gap the passing interrupt test had hidden — nothing had exercised the turn *after* an interrupt, which is the bounded nudge itself
- Designed completion around independent evaluation rather than implementer confidence, with design and security gates that stopped two of three first attempts correctly, and recorded the defect every gate still passed — a job-level permissions block silently replacing the workflow-level one — caught by a dispatched run thirty seconds after merge
- Carried agent execution policy in the repository as durable versioned artifacts rather than per-session steering, stating rules the factory enforces: no self-certification, every lane has a clock, an engine review verdict is supplementary rather than completion evidence, and durable policy with disposable workers