# Building sheal

*March 2026*

It started with a pattern we kept seeing: the same mistakes, session after session.

A Claude Code session would fail on an assumption or file path. We'd fix it. Next session, same mistake. Same wasted tokens, same frustration, same sneaky bugs and drift popping up in the codebase. Remembering to add to AGENTS.md manually, so that codex, amp, opencode would also catch it. Remembering to maintain it.

The problem isn't the AI itself, the problem is that sessions are stateless. Every session starts from zero. There's no memory, no learning and no feedback loop. You can try to fix it with a memory CLI or MCP, of course. But how do you know what the agent is actually learning and extracting?

sheal stands for "self healing" and is our answer to close the loop.

## How it works

```
Session Capture (Entire.io / Claude Code / codex / gemini / amp)
    |  session transcripts, diffs, metadata
    v
Self-Healing Engine (sheal)
    |  failure patterns, learnings, rules
    v
Human in the Loop (HIL)
    |
    v
Agent Configuration (CLAUDE.md, AGENTS.md, etc.)
    |  improved behavior
    v
Next Session (fewer mistakes)
```

After 20 sessions with sheal running, repeated mistakes dropped to near zero. Token waste fell. The agent got smarter because the feedback loop was closed. The human got less frustrated.

Now it is your turn. Help us build the self-healing loop for coding agents. What will you bring?

---

*— Luisa*