Reliability · 2026 · Solo — architecture, build, ops

Self-Improvement Loop

A daily loop where agents research their own improvements and safely apply them with snapshot + rollback.

The problem

Agent prompts rot. Manually tuning 11+ agents doesn't scale. I wanted the system to improve itself — without letting a bad edit silently break production.

Architecture

Key decisions

01

A human-authored safety envelope

Snapshot before every write, roll back on failed verification. The loop can experiment freely because it structurally can't do lasting harm.

02

Separate research from apply

Proposal (cloud) and mutation (local) are different trust levels, so a runaway proposal can't reach production on its own.

03

A durable Notion queue

Improvements are reviewable artifacts, not opaque in-memory decisions — I can see and veto what the loop wants to change.

Metrics

27
improvements auto-applied
4
auto-rollbacks caught
0 edits
manual prompt edits this cycle