Self-Improving Agent Ecosystem

Self-Improving Agent Ecosystem is an open-source reference architecture created by David Zhang (GitHub: Git-on-my-level) and released under the MIT License. It provides the contracts, schemas, skills, examples, and deterministic helpers needed to build evaluator-driven systems that can safely observe, propose, test, promote, and learn — without conflating activity with genuine improvement. The repository is intentionally not a ready-to-run autonomous agent; it is a structured foundation that future agents build upon.

What It Is

Self-Improving Agent Ecosystem is a reference kit for constructing safe, evaluator-driven feedback loops in agentic AI systems. It addresses a core challenge in autonomous systems: ensuring that a system's self-modification cycle is grounded in verifiable evidence rather than unchecked automation. The kit supplies the architectural contracts, JSON schemas, portable agent skills, and small Python helpers that any self-improving loop would otherwise have to rediscover independently.

Three Loop Archetypes

The repository defines three composable loop archetypes that share the same underlying planes:

• optimize — ranks candidates against a scalar objective behind a hard correctness gate, using an AVO-lite style ratchet to prevent regression.
• maintain — drains a work queue (issues, pull requests, dependency updates, alerts) through triage, candidate, review, publish, merge, and outcome check stages, under a WIP limit and maximum item age. The README recommends most deployments start here.
• observe → propose — sensors convert telemetry, usage traces, transcripts, and human decisions into proposals only, never code. This archetype feeds the maintain queue, and a maintain loop can run optimize on any item that has a scorer.

Four-Plane Architecture

The reference architecture separates concerns into four deliberately distinct planes:

• Experiment kernel — AVO-lite provides disposable worktrees, hard scoring, verification hooks, accepted-only lineage, an append-only ledger, and stagnation handling.
• Workload control plane — domain-owned sensors, evaluators, promotion, deployment, outcome checks, and rollback.
• Authority and distribution — reviewed policy and portable agent skills; no prompt silently grants deployment or credential authority.
• External observation — a separate process or host checks loop freshness, lineage gaps, observer blindness, and delivery health.

Key design principles baked into the architecture include: correctness and quality are treated as separate concerns; candidates are isolated from canonical state; a would-be winner is adversarially verified; every attempt retains exact evidence and lineage; and human decisions are labelled evidence with autonomy earned per class of action.

Setup Path

Requirements are minimal: Git, Python 3.10+, and a POSIX shell. The validation helpers use only the Python standard library. After cloning, an optional script fetches AVO-lite source without executing it, and init-ecosystem.sh creates a local scaffold from public templates. Three scaffold profiles — local, project, and live — share one architecture but scale operational requirements to consequence. Every scaffold includes a MISSION.md so outcome, acceptance evidence, authority, non-goals, and stop conditions are explicit before any automation runs.

Explicit Non-Goals

The README is direct about what the kit does not provide: a universal objective function, a fleet-wide write agent, an excuse to give an LLM production credentials, a replacement for domain tests or incident response, or a claim that worktree isolation constitutes a security sandbox. This scoping makes the kit a disciplined foundation rather than an overreaching framework.

Current Status

The repository was created in late August 2026 and last updated in September 2026, with 40 stars and 1 fork at time of indexing. It is actively maintained on the main branch and includes contract tests and reward-hacking regression fixtures alongside the core architecture documentation.