Dev shares safety architecture for self-improving agent: out-of-band release controller with passkey approvals
viksit · x · 2026-11-01
Developer viksit details the safety architecture of Myra, a self-improving agent built on pi-durable, tackling how to guarantee human approval for code changes.
- Isolated candidates: self-updates are prepared in an isolated workspace, checks run, and the diff/results/preview shown to the human
- Passkey approvals: release is approved via a separate review page with a passkey; the approval is bound to the exact code and build artifacts — any change requires re-approval
- Key design: the release controller lives outside the application Myra can edit; it owns approval, release checks, and rollback. Myra has no tool call to approve its own changes, and updating the controller requires separate human maintenance
- Health checks: after approval the controller installs and verifies the version, rolling back on failed startup
Next step: an observer model that mines usage traces to suggest improvements and prioritize features.
More from AGI Musings
- Reddit debate: cloud LLMs change rules on a whim — only local models are reliable — Substantial_Swan_144 · 2026-11-01
- "2 years to escape the underclass" was AI industry fear marketing to boost valuations — Signalman23 · 2026-11-01
- Ben Goertzel: the common framing of the 'bitter lesson' is 'quite stupid' — bengoertzel · 2026-11-01
- Can LLMs turn surprising observations into research questions and discover physical laws? — rohanpaul_ai · 2026-11-01
- Gamers push back: 'We want MORE AI in gaming, not less' — petewoodbridge · 2026-11-01
- 'Don't recreate existing games': AI platform terms spark open-asset fears — TinfoilTricorn · 2026-11-01