Agent Drift Circuit Breaker

Hit the brakes before your multi-agent setup wrecks the repo

Problem

People report that LLMs over-defend themselves on complex engineering tasks, producing rigid hardcoded logic and extreme regex checks that run slowly and fall apart in multi-agent and low-level call chains. Meanwhile AI-generated PRs flood repos like a DDoS, overwhelming maintainers. The real problem isn't model intelligence, it's that nobody watches when an agent's behavior starts drifting from the original goal.

Solution

A lightweight daemon that sits in your local env or CI, reads agent traces and diffs, and scores each step with configurable rules plus a small model to judge intent drift. When it detects over-defensive rewrites, unrelated file creep, or abnormal commit frequency, it pauses that agent, snapshots the state, and pings you for confirmation instead of letting it trash the repo.

Users

Small teams and solo devs running multiple coding agents, especially those with CI gates but no agent behavior audit

MVP

Only support trace reading for Claude Code and Codex; ship 5 heuristic rules (unrelated file edits, regex stacking, per-PR commit cap, hardcoded constant ratio, cross-module call explosion); pause locally and prompt on anomaly; one-command hook into pre-commit or CI; generate a one-page daily drift report

Revenue

One-time purchase for personal use, per-repo subscription for teams, paid add-on for CI integration; later sell custom rule packs

Why now

Agentic coding is moving from single scripts to multi-agent collaboration, so failure cost jumps from rewriting a snippet to polluting a whole repo. Rules plus small models are enough to build this, and one person can ship it in weeks without training a foundation model