A production pattern for autonomous, self-improving multi-agent operations loops built on top of an LLM coding agent (Claude Code). It is the genericized core of a real system I run daily: a long-lived "conductor" loop that orchestrates fan-out subagents, verifies its own outbound actions, distills failures into a self-improving learnings ledger, and watches structured data sources by polling their JSON APIs directly instead of scraping.
I build these because the interesting problem in applied AI isn't the model call - it's the harness: how do you get a fleet of agents to do real, side-effectful work reliably, safely, and without a human babysitting every step? This repo is the distilled answer.
- Multi-agent orchestration - deterministic fan-out / pipeline control flow over subagents, with
barriers only where cross-agent state genuinely requires them (
examples/orchestrator.py). - A conductor / tick loop - one long-lived loop that re-orients off sources of truth each tick,
runs phased work, self-reschedules, and stays quiet unless a human decision is needed
(
docs/conductor-pattern.md). - Self-improvement - every failure or surprise becomes a durable, append-only learning with a
trigger, a rule, and an automated check, so the system stops repeating its own mistakes
(
docs/self-improving-loop.md). - Postflight verification - no action is reported "done" until its real-world effect is confirmed (the banner appeared, the message bubble rendered) - the difference between "the model said it worked" and it actually worked.
- Structured-source polling - watch dozens of sources by hitting their JSON endpoints directly
(
examples/structured_source_poller.py), because aggregator/search snippets are ~60% stale. - Human-in-the-loop by design - the loop executes autonomously but escalates the irreducible human steps (an approval, an anti-bot CAPTCHA, a credential) through a single notification channel, batched - never silently blocked, never spamming.
These are the exact problems you hit deploying agentic systems for real users: ambiguity handling, reliability under partial failure, verification, safety rails, and knowing which 5% must stay human. The model is a commodity; the translation layer - turning a fuzzy real-world goal into a harness that ships reliable actions and measures itself - is the part that's still rare. I've also run this pattern as an internal-AI-adoption program inside an engineering org, with a measured outcome of ~30 minutes saved per engineer per day.
docs/conductor-pattern.md # the long-lived tick loop (orient -> phases -> verify -> reflect -> reschedule)
docs/self-improving-loop.md # capture -> apply -> verify -> reflect; the learnings ledger
examples/orchestrator.py # fan-out / pipeline subagent control flow
examples/structured_source_poller.py # poll JSON APIs directly, score, alert
Patterns extracted from a system in daily production use. Generic by design - no domain specifics, no credentials, no private data. MIT licensed; use the patterns freely.