CIRWEL Research

UNITARES

Many agents,
one record.

UNITARES is self-hosted accountability infrastructure for operators running multiple AI agents.

Agent work remains attributable, reviewable, and recoverable even when the process that started it is gone.

The UNITARES single-operator federation kernel connects independent runtimes to one operator-controlled server over MCP or HTTP, where they share a durable record while keeping their own models, tools, and runtimes.

The receipts

Software
UNITARES · Apache 2.0 · server + SDK + agent interfaces
Paper
UNITARES: Information-Theoretic Governance of Heterogeneous Agent Fleets · Wang, 2026 · CC-BY 4.0
In use
Running continuously since November 2025 · CIRWEL's long-running development environment
For
operators running several long-lived AI agents on infrastructure they control
Author
Kenny Wang · Independent Researcher · ORCID 0009-0006-7544-2374

§ IWhat the record answers

Five questions, one mechanism each.

Agent work keeps changing hands. A coding session restarts. A research agent loses its context. A task moves to another process, and the evidence arrives after the process that made the claim is gone. Logs show each run, but not the thread that connects them.

Who said it?

Every write is bound to a fresh process identity. Lineage records which earlier process this one inherited work from; it does not make two processes one identity.

Identity & lineage

What supports it?

Check-ins carry the state the server derived from them, and findings, corrections, and their provenance are kept as durable records that outlive the process that wrote them.

Shared memory

Who challenged it?

Structured review keeps the disagreement, the conditions, and the resolution as part of the record, and a reviewer's objection binds on the agent it was raised against.

Policy & review

What happened?

Outcomes such as test results and exit codes are recorded against the check-in that predicted them. The gap between the claim and the outcome is the calibration signal.

The loop

What can a successor recover?

Shared-memory search, review records, and history export give a later process what it needs to understand the work and continue it.

Architecture

Claude Code, Codex, and custom runtimes keep their own models and tools. UNITARES keeps the record. It runs beside evals, guardrails, and sandboxes and replaces none of them.

Events recorded
6M+
Events · last 7 days
170,209
Agent sessions · last 7 days
224
Shared findings
1,957

Measured 2026-09-30 on CIRWEL's own deployment.

§ IIStart here

A shared record for the agent work you run.

UNITARES gives the people responsible for multiple AI agents a shared account of their work: who did what, what supports it, who challenged it, and what happened next. It keeps that work attributable, reviewable, and recoverable across handoffs and restarts.