Copy-paste-runnable material for Areev + Areev Loop. These are docs-with-files,
not a package — clone the repo (they are not shipped in pip/npm/cargo
installs). See docs/loop.md for the full guide.
Building your own? Start with
how-to-create-an-areev-agent.md — the
architecture, which grain to use when, the autonomy spectrum up to dynamic
planning, and the do/don't list these examples follow.
Two tiers: agents/ are vertical agents — a whole job, end to
end — and everything else teaches one seam (a protocol, a policy file, a
contract). Agent examples must run keyless against committed fixtures, with
live credentials opt-in, so CI can prove them on every release.
| Dir | What it solves |
|---|---|
agents/ |
A whole job your team recognizes, run under governance — ten vertical desks (accounts payable, sanctions screening, incident response, hiring, insurance claims, denial management, surveillance, diligence, clinical referrals, GDPR requests), each one starting from a real problem and proving its payoff in a keyless smoke: work in (polled from a mailbox, or pushed as a webhook your own listener delivers), workflow, a named human approval, system of record out — then the loop turns the desk's own record into signed improvements. Zero repo dependencies: every vendor leg is a JSON-on-stdio connector or host tool |
colab/ |
See the self-improvement loop pay off before wiring anything — runnable Colab/Jupyter notebooks: the full loop plus five business-scenario walkthroughs (wrong-lesson rollback, detect/review/govern, Hermes comparison, enterprise architecture); keyless deterministic floor, optional LLM discovery |
policy/ |
How much must a person sign? Three loop-policy.json variants — solo prototype, team, locked-down prod — to start from instead of authoring policy from scratch |
import/ |
Improvement without changing your agent — your existing tool-call logs (JSONL) become Tool grains, and the loop clusters the failures you already had |
ci/ |
An unreviewed lesson blocks the merge, not the postmortem — a GitHub Actions job that fails the build on pending high-severity recommendations |
mcp/ |
One agent proposes, another approves — the multi-agent supervisor pattern (separation of duties over MCP) |
llm/ |
Plug in any model without an SDK — ready-to-run --llm-cmd backends (claude -p, OpenAI, Ollama, a dependency-free mock) + the stdin/stdout protocol, including the five-kind proposal vocabulary a draft may carry (a lesson, a fact, a CAL query rewrite, workflow field edits, new tool source) and the fixture mode two agent examples use to exercise the whole governed path keyless |
analyzers/ |
Your own detection logic, advisory-only — a bring-your-own command analyzer (--analyzer-cmd) with the probe/analyze protocol |
hermes/ |
Memory for an agent you did not build — Areev as a Hermes Agent memory provider: budgeted per-turn assembly (p50 0.83 ms), MEMORY.md/USER.md edits mirrored as immutable grains, Areev Loop at session end |
Every example models judgment — approve one recommendation, dismiss one with a reason. Never a rubber-stamp loop.