Instead of making the model reliable, he gated every risky action around it
thebvg · reddit · 2026-07-27
The author stopped trying to make the model itself reliable and instead put reliability into the layers around it.
- In a multi-tenant product where the model writes most of the code, prompt edits only helped marginally.
- The real fix was deterministic gating: a hook typechecks every file write, static rules block forbidden patterns outside an allowed package, and an allowlist controls when the agent can push or open a PR.
- The model is treated as an unreliable component inside a normal system.
- The goal is not to make its mistakes disappear, but to make them cheap, visible, and contained.
- The author still asks for better ways to test the non-deterministic part itself.
More from coding & agent
- AI agents may lose to UIs on repetitive work, but win on novel tasks — bendee983 · 2026-07-27
- GPT-5.6 Sol and Terra are far more efficient, but Codex is the bottleneck — haider1 · 2026-07-27
- NousResearch update keeps five-agent streaming at 60 fps and makes switching instant — max_paperclips · 2026-07-27
- An Opus subagent swarm reportedly kept grinding for 7 hours straight — basedalexandoor · 2026-07-27
- AgentHost unifies Claude, Codex and local models in one agent control room — Stevekaplanai · 2026-07-27
- Agent Context Lens v0.2.0 adds local inspection of Codex AGENTS.md chains — Cicerron · 2026-07-27