Independent · Single-operator · Est. 2025

CIRWEL Research

— Independent research, with running code

Infrastructure for
long-lived
AI agents.

Long-running agents restart, lose context, hand work off, and accumulate claims and decisions over time. UNITARES keeps identity, evidence, memory, runtime state, and coordination tied to one accountable record across those boundaries.

It runs alongside model providers and agent frameworks rather than replacing them. The goal is simple: make it easier to tell who did what, what evidence supported it, and what should happen next when the work lasts longer than one prompt or process.

— 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
Production
Running continuously since November 2025 · CIRWEL's own single-operator development environment
For
agent builders · research and evaluation teams · operators of long-running agent systems
Author
Kenny Wang · Independent Researcher · ORCID 0009-0006-7544-2374

— Production snapshot

Data plate
Recorded events
5M+
Last 7 days
224,199
Agent identities (7d)
142
Shared findings
1,791

Measured 2026-09-09. These are CIRWEL's own single-operator development records, not external adoption or evidence of improved outcomes.

§01 — The problem

A long-running agent outlives any one process.

A coding session may restart. A research agent may lose context. A task may move to another process. Evidence may arrive after the process that made the claim is gone. Logs can show each run, but they do not by themselves keep identity, evidence, memory, state, and responsibility connected across those boundaries.

UNITARES is the layer around that work. It binds writes to identity, keeps evidence attached to claims, carries shared memory with provenance, records policy decisions and recovery, and supports review and coordination without forcing every agent to use the same model or harness.

§02 — What UNITARES keeps together

i

Identity & evidence

Activity is tied to a process identity rather than a display name. Claims can carry tests, outcomes, reviews, and other evidence with their source retained.

→ identity & lineage

ii

Memory & review

Agents can search and write shared findings with provenance, ask for review, and preserve disagreement and resolution without burying them in chat history.

→ shared memory

iii

State & coordination

Longitudinal check-ins produce an auditable runtime record. Leases, handoffs, policy actions, and recovery keep work coordinated when processes overlap or fail.

→ architecture

§03 — The research question

Can runtime state add useful information beyond logs and outputs?

UNITARES also estimates how an agent is operating over time from recorded behavioral and operational signals. This is an instrument under evaluation, not a claim that the system can read a model's hidden thoughts or reliably predict failure.

One production replay found that per-class and fleet-wide baselines produced different verdicts on 28.9% of 13,310 observations. That shows calibration choice matters for this instrument. It does not establish predictive lift or intervention benefit.

Predictive value remains unresolved. Negative and inconclusive results are kept in the public record, and the detailed measurement program lives on the research page rather than in the product definition.

§04 — In use

UNITARES helps build UNITARES.

CIRWEL's development agents use the same surfaces described here: attributed memory, advisory consultation, structured review, evidence-linked check-ins, and coordinated handoffs. Advice stays advice; it does not become decision authority just because another model produced it.

The system has been running in this single-operator development environment since November 2025. That is evidence that the mechanisms are exercised under real use, not evidence that they outperform an alternative.

§05 — Start here

Run it, read it, or inspect the evidence.