zouroboros
A multi-agent orchestration platform with persistent memory and an autonomous self-improvement loop, built natively for Zo Computer (by marlandoj; MIT; TypeScript; v2.1.0, June 2026; ~19★, 757+ tests, 8 npm packages). A single repo that fuses four of this spoke’s threads at once — orchestration, memory, self-improvement, and resilience — which makes it, like hermes-agent and gstack, a convergence instance rather than a single-feature tool.
What it does
- Orchestration — routes a DAG of tasks across five executors (Claude Code, Gemini, Codex, Hermes, Mimir) via 8-signal adaptive routing (capability, health, complexity, history, procedure, temporal, budget, role), with retry/fallback cascades. A 57-role registry maps MetaGPT-style SOPs onto the swarm.
- Hybrid memory — SQLite + vector embeddings across episodic, procedural, and cognitive layers, with domain-context injection and RAPTOR hierarchical retrieval over 5 RAG surfaces; soft-degrades to text-only search when no embedding key is present.
- Self-enhancement — daily introspection cycles (a 4-watcher “Health Council” over runtime / orchestration / capability / knowledge) audit capability gaps, prescribe fixes, and evolve procedures through 8 playbooks.
The notable mechanism — a consensus verification gate
What distinguishes its self-improvement loop is the feedback signal: procedure changes don’t ship on one model’s say-so — a “Consensus Gate” routes each change through three frontier models in parallel (GLM-5.1, Kimi-K2.6, MiniMax-M2.5) and requires agreement. That is the spoke’s “the bottleneck is verification, not generation” thesis answered with multi-model agreement as the truthful-enough signal — a different design point from autonovel‘s mechanical-scan-plus-LLM-judge or autoresearch‘s objective metric. The recursion is fenced by guardrails: category-aware circuit breakers (8 failure types), 5-layer loop guards against infinite recursion, and stagnation-triggered auto-recovery.
Tier
T1 — first-party project repo: code + 757 tests + 8 published packages, a concrete artifact, the
spoke’s convention for project source. Weaknesses recorded: young and small (~19★, single author), and
platform-locked — “built natively on Zo Computer, not portable” is a stated design
choice, so it’s an instance of the patterns, not a general framework. freshness: volatile (active repo).
Related
agent-orchestration · agent-memory · self-improving-agents · loop-engineering · agent-loops-verification · agent-guardrails · durable-agents · hermes-agent · zo-computer · agentic-tooling