Skip to content
View cirwel's full-sized avatar
🚞
🚞

Block or report cirwel

Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
CIRWEL/README.md

CIRWEL stack — runtime governance for heterogeneous AI-agent fleets

Kenny Wang — CIRWEL Systems

I build runtime safety infrastructure for autonomous AI agents — the layer that operates after deployment, while agents are actually running. Agents fail gradually before they fail visibly: drifting, thrashing, growing overconfident on stale context. CIRWEL builds the state layer that lets an agent — and its operator — notice and act on that drift before it becomes an incident.

The stack has run continuously on a single-operator development fleet since November 2025. That's a stress test and a telemetry corpus, not a claim of external adoption — external validation is the next step.

cirwel.github.io — the full index

Papers, systems, datasets, and decks, all in one place. That page is canonical; this profile is just the front door.

Index UNITARES Paper v6


The work, in four lines

UNITARES Governance runtime — MCP + HTTP, Postgres-backed Agents check in; it grades drift and calibration against each agent's own baseline and returns a verdict (proceed / guide / pause / reject) every call. Live since Nov 2025. → repo
Anima Physical longitudinal testbed Raspberry Pi 4 + sensor stack mapping real temperature, light, humidity, pressure, and system telemetry into EISV trajectories. The source behind the papers and dataset; the richer creature/art interface lives in the repo. → anima-mcp
Research 3 papers / preprints Information-theoretic fleet governance (v6, DOI) · trajectory identity (Wang 2026b) · digital proprioception (Wang 2026c).
Datasets Published telemetry corpora 32,181 labeled EISV trajectories (20,655 real) · verdict-counterfactual repro kit.

Start here: unitaresdocker compose up -d --wait && make demo drives a synthetic agent through seven check-ins and prints the verdict at each step.

For reviewers

  • What this is: runtime state telemetry for agent fleets after deployment — the layer between evals/guardrails and incident response.
  • What this is not: not an output filter, not a sandbox, not an ethics oracle, and not yet a claim of external adoption.
  • Current ask: external pilots and design partners who already run autonomous agents long enough for drift, calibration, and recovery to matter.

Full index ↗ · GitHub · HuggingFace · ORCID · CIRWEL Systems · founder@cirwel.org

Pinned Loading

  1. unitares unitares Public

    Runtime state telemetry for long-lived AI agents — they check in, UNITARES grades drift and calibration against the agent's own baseline, returns a verdict each call. Live deployment since Nov 2025…

    Python 3

  2. unitares-paper-v6 unitares-paper-v6 Public

    UNITARES v6 paper: Information-Theoretic Governance of Heterogeneous Agent Fleets

    TeX

  3. unitares-host-adapter unitares-host-adapter Public

    Thin client bindings that mount UNITARES governance into agent hosts (Hermes, Claude Code, Goose, ...)

    Python

  4. unitares-governance-plugin unitares-governance-plugin Public

    Claude Code / Codex plugin that mounts agents into UNITARES governance — check-ins, dialectic, runtime verdicts.

    Python

  5. anima-mcp anima-mcp Public

    Raspberry Pi running UNITARES — longitudinal testbed cited in the papers.

    Python 1

  6. eisv-lumen eisv-lumen Public

    Governance benchmark dataset — 21K agent-state trajectories with verdicts (HuggingFace).

    Python 1