CIRWEL

CIRWEL is an independent research and systems lab, founded by Kenny Wang, building runtime governance, continuity, and observability infrastructure for autonomous AI-agent fleets.

Agents fail gradually before they fail visibly — drifting, thrashing, growing overconfident on stale context. CIRWEL builds the state layer that lets agents notice and act on that drift before it becomes an incident.

agent acts  →  check-in  →  calibrated state + verdict  →  self-regulates  →  audit trail

The CIRWEL stack has run continuously since November 2025 as a single-operator deployment. That is a stress test, not a claim of external adoption.

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 a universal ethics oracle, and not yet a claim of external adoption.
• Initial wedge: teams running long-lived coding, research, or operations agents add one MCP/REST check-in per unit of work plus outcome events from tests, tools, and other hard signals.
• Current ask: external pilots and design partners who already operate autonomous agents long enough for drift, calibration, and recovery to matter.

Read order:

1. 3 minutes: unitares → run docker compose up + make demo.
2. 10 minutes: add unitares-governance-plugin, unitares-host-adapter, and anima-mcp.
3. 30 minutes: read unitares-paper-v6, reproduce §11.6 with unitares-repro-v6, and inspect eisv-lumen.

---

Stack

Runtime governance UNITARES is the MCP + HTTP governance server. Agents check in; UNITARES tracks a live state vector per agent and returns a verdict — proceed, guide, pause, or reject — so agents self-regulate before circuit breakers fire.	Host integrations Governance plugin, host adapter, and hermes-agent mount governance into Claude Code, Codex, and other agent hosts.
Peer review on disagree When an agent's confidence and the system's assessment diverge, UNITARES runs a short dialectic with peer agents — or an LLM, when no peers are around — before anything halts. Training data: dialectic-dataset.	Research and observability Paper v6 (concept DOI 10.5281/zenodo.19647159) on heterogeneous-fleet calibration · repro kit for the §11.6 verdict counterfactual · Discord bridge for live operator visibility.

Featured Repositories

UNITARES · Governance plugin · host adapter · Anima/Lumen · EISV-Lumen · Paper v6 · repro kit

---

HuggingFace · ORCID · CIRWEL Systems · founder@cirwel.org