pexels ai25studioai 5475809

Blume: The Sidecar That Watches Your Coding Agents and Fixes Their Rulebook

Every harness on this list either runs a model or orchestrates the ones you already run. Blume does neither. It’s a sidecar – a narrow tool that watches the harnesses you’re already using and keeps the hidden files behind them (your CLAUDE.md, your rules, your skills, your hooks) honest.

The problem it targets

Anyone who’s used a coding agent for more than a few weeks knows the pattern: the agent misunderstands the same thing three times, you re-explain the same rule in three different chats, and none of that friction ever makes it into your project’s actual configuration. Each instance is buried in a different conversation, so the pattern is invisible until you happen to notice it.

How the pipeline works

Blume runs a small internal chain of agents against your own conversation history: an extraction agent reads each transcript and pulls out “signals” – friction points with evidence and a pain score; a clustering agent groups new signals with similar ones from past conversations rather than treating every one-off as noise; a promotion step decides, from local thresholds, when a cluster has enough evidence (signals × conversations × pain) to act on; and a planning agent turns a promoted cluster into a concrete operation – create, update, or remove a rule, hook, or skill file – complete with a diff. Nothing gets applied without your review: you get the evidence and the exact diff, and you apply, dismiss, or save it for later with one-click undo.

Local by design, and usage-aware

The analysis itself runs through your own agent CLI – Claude Code, Codex, or Cursor – using your existing plan’s quota, and your conversation history is stored on your device rather than uploaded anywhere. Beyond the improvement loop, Blume’s dashboard doubles as a control room for the harnesses you already run: it tracks how much of each provider’s usage you have left before you hit a wall, and shows what every agent is doing right now, including the exact moment one finishes or needs your approval.

Why it belongs in a harness roundup

Blume is a useful reminder that “harness” isn’t just the thing that runs the model – it’s increasingly the layer of tooling that makes the harness you already chose better over time. If you’re running Claude Code or Cursor daily and keep hitting the same wall with your project’s instructions, Blume is a narrower, lower-commitment thing to try than switching harnesses entirely.


Leave a Reply