LoopGuard convergence monitor for coding agents

Circuit breaker and convergence checks for runaway AI coding loops

Problem

Developers describe Ralph-style agent loops getting stuck in local minima, with stop conditions impossible to define for non-deterministic tasks like complex GUIs. Another report shows Claude burning 30-70% of a 5-hour limit in 10 minutes, and users can't tell whether the loop is progressing or where the quota went.

Solution

A thin CLI/proxy that wraps coding agents such as Claude Code and OpenCode: it fingerprints tool calls and file diffs to detect repeated actions, no-progress edits, and context bloat. When thresholds trip it forces a checkpoint, asks the agent to change strategy, or escalates to a human, and produces a per-task convergence and token report.

Users

Individual developers and small-team tech leads who run AI coding agents daily

MVP

Wrap the agent command and capture step logs; detect loops via file-diff fingerprints and tool-call sequences; configurable thresholds and checkpoint prompts; per-task token and round accounting with a report

Revenue

Subscription with a single-project free tier; Pro adds multi-project, team dashboards and alerts; optional per-agent-hour billing

Why now

Coding agents are now mainstream but convergence and cost control are fresh, unsolved pain; a local-first proxy is feasible for a solo builder in weeks