self-improving-agent-looplisted

# Building an agent loop that actually improves Distilled from a working system that monitors three social platforms, proposes changes, and grades its own past advice against measured outcomes. **The default failure mode of every long-running agent:** it produces advice, nobody records whether the advice worked, the knowledge base fills with plausible claims, and within a month every recommendation is confidently wrong. Everything below is defence against that. --- ## 1. Specialists that sleep, not one generalist One agent per job. Each is **woken only by the event it owns** and does nothing the rest of the time. | Agent | Woken by | Owns | |---|---|---| | Design | an asset needs judging/replacing | the visual artifact, scored against the account's own baseline | | Copy | the words are the problem | titles, hooks, first 15 seconds | | Analyst | daily | one briefing: what changed, why, **one** instruction | | Platform (per surface) | that surface's metrics move | platform-specific behaviour | | Strategy | weekly, or on a trigger | what to make next | | Critic | weekly | grading the others, deleting unsupported claims | Why specialists win here: - **Cost.** Every agent's context is paid for on every run. A generalist that reads everything is expensive *and* vague. - **Scoring.** A narrow agent can be scored on a specific measured number. A generalist cannot. - **Currency.** Each specialist can have its own small research beat keeping it up to date, without polluting