Agentic speed needs a proof system.

Day 05 · 2026-10-02 · LinkedIn

Short video caption

AGENTS.md, CLAUDE.md and skills can describe standards; real gates live at tool, identity and CI boundaries. Start in a worktree. A Python PreToolUse hook can deny covered writes outside its root; use OS sandboxing or deny arbitrary shell. Log consequential tools to an external WAL. CI rejects zero tests, stale reports and a surviving deliberate fault. See the accompanying tested harness example. This is an architecture recommendation, not a claim that one hook covers every product route. #EricFieldNotes

Full written post / accessible read

A good coding agent can move from spec to patch to tests in minutes. That is a real advantage, if the release gate can tell a true pass from a performance. In this run the report is green, but zero relevant tests ran and it belongs to the previous checkout.

The reviewer now inherits a wrong tree, stale report and skipped suite. AGENTS and CLAUDE files can state a rule, but cannot validate that evidence. The change can merge with no test that would catch its bug. The cost is false confidence in the release gate.

Start each run in a separate Git worktree with a shell launcher. Give a trusted Python pre-tool hook the allowed root; resolve a proposed file path and exit two when it escapes. Deny arbitrary shell or run it in an operating-system sandbox. After-tool hooks append a write-ahead receipt, but cannot undo the action.

Then let CI, outside the agent, verify the expected worktree, a current report, a nonzero test count, and one deliberately broken behavior that turns the suite red. Compare the business outcome with an independent oracle. Do this because instructions express intent, scripts enforce covered calls, and CI decides whether the evidence is real.

#EricFieldNotes

Runnable agent-control example: worktree, hooks, WAL and CI

Evidence and boundary

On-screen boundary: ILLUSTRATIVE HARNESS · NOT PRODUCT TEST. The sources below support documented mechanisms and specifications; illustrative scenarios are not presented as measured incidents.