James Washington
writing about archive utils free game ideas replays
Automation

agent-merge: Jev-gated self-merge for agent branches

The one sanctioned way an AI may merge its own work: a script that reviews it first and refuses if it didn't show its work.

built Oct 2026 updated Oct 2026 Running a team of AI agents

how it works

Two layers, fail-closed: mechanical checks (agent/* only, clean tree, merges cleanly, CI green, never-stage paths, size cap, opt-out file, due-diligence note with Verification) then seven yes/no TypeSafe Jev classifier questions with thresholds; Claude Code autoMode allow-rule carves out only this exact invocation.

Automation: Something that runs on a schedule by itself.

The code for this one lives in a private repo. This page explains the idea so you can build your own.

auto-modeclassifiergitguardrailmulti-agent

related

  • Agent git fleet: worktrees, sweeper, ship Automation Gives every AI session its own copy of the code, then automatically opens reviews for finished work and cleans up.
  • Git discipline for agent fleets AI skill Rules that stop many AI assistants working in the same code from trampling each other's changes.
  • Run a classifier in shadow mode before it can gate Technique Let a new AI judge watch silently until there's evidence it's right.
  • Agent Scratchpad: a moderated GET-addressable commons for sandboxed LLM agents Note An idea for a public message board that locked-down AI agents could use to coordinate, built so the humans can always watch.
  • Loop engineering (Addy Osmani clipping) Note A saved article about making AI agents prompt themselves in loops instead of a human prompting them each turn.
← everything I've built
© 2026 James Washington
github twitter rss