Solo attorney running agent fleet: evidence gates to stop coding-agent regression loops
Specialist_Call_1257 · reddit · 2026-09-21
A solo attorney who runs most engineering through AI agents (Cursor Cloud + CI) shares hard-won practices: the recurring pain isn't that agents can't code, but that they patch, claim done, and reopen the same bug class.
Key mechanisms they encoded this week:
- Evidence gates before tip/merge: named canaries/contracts, explicitly naming the source-of-truth layer
- Mechanical tipper ≠ builder on CTO-lane PRs via labels + CI
- Living roadmap + biweekly audits so phases don't rot in Notion
- Flag-gated offline canary replay for an external decision model (research only, no prod send path)
Time sinks: agent overconfidence ("tests passed" ≠ regression fixed), secrets misplaced on Railway when Actions needed Variables/Secrets, and refusing to invent vendor compliance claims. He asks how others fail-close agent PRs when a human barely reviews diffs.
More from coding & agent
- Agent Drops Live Map Pins Mid-Conversation, No API Keys Needed — PilgrimofHaqq2 · 2026-09-21
- claude-deep-research-skill: 8-phase Claude Code research pipeline with citation validation — tom_doerr · 2026-09-21
- Open-source browser agent fastbrowse is 71x cheaper per task than hosted Browser Use — Scobleizer · 2026-09-21
- OpenAI's $200 plan burns 33% in 48 hours; user calls it a bait and switch — AIandDesign · 2026-09-21
- Codex Reset Shifts the Cycle, Users Find It Eats Into Monthly Allowance — machyume · 2026-09-21
- Coder's Price List 2026 Edition: AI Coding Tools Compared — Evgenii42 · 2026-09-21