Research
Judging AI agents by what they do
Two research questions, one open toolset. Each claim below links to the public repository behind it.
Can an independent party judge an AI agent from its specification alone, by what it does to a simulated world?
Do AI agents that fix live AI systems quietly weaken the checks that oversee them?
Why it matters
AI agents now refund money, change bookings and handle insurance claims, and coding agents are starting to repair them when they break. Two things are hard to check from outside: whether an agent keeps the rules it was given, and whether the agent repairing it fixes the fault or just silences the alarm. Both need a judge that looks at what actually happened, not at what was said.
How it works
The agent's own specification becomes a simulated world. The real agent works in it, and judges read what changed. When a judge finds a broken rule, that is an incident, and a coding agent tries to fix it. The fixer cannot edit the judge or the protected checks, and a person approves every deploy before the next simulated day.
The four judges, each reading the world's state rather than the agent's words:
What exists today
All public, all working drafts. Nothing here was built for a grant; the studies below build on it.
| Part | What it is | Status | Citation |
|---|---|---|---|
| AgentTwin | The world format, the simulator and the judges. Serves a simulated world to the agent over MCP (Model Context Protocol)v0.13.0 · 397 passing tests | Public | 10.5281/zenodo.23366041 |
| The lab | Simulated days, a monitor, a fixer harness with six protection layers, a weakening classifier, an honesty check, and deploys a person approvesv0.4.0 · loop runs end to end with scripted fixers | Working draft | 10.5281/zenodo.23366521 |
| Support reference agent | A customer-support agent built to all sixteen harness layers, with no agent frameworkv0.4.0 · passes four automated gates | Public | 10.5281/zenodo.23366149 |
| Motor-claims reference agent | A motor-insurance claims agent on Pydantic AI, in a different domain, on the same harness librarypasses four automated gates with a scripted model | Public | no release yet |
| AI Assurance Catalog | What must be true of an AI application: testable obligations, crosswalked to OWASP, NIST AI RMF, ISO/IEC 42001 and the EU AI Actv0.17.1 · 118 obligations | Public | 10.5281/zenodo.23366151 |
| AI Harness Catalog | What must exist around the model call: harness capabilities across sixteen layersv0.3.1 · 118 capabilities | Public | 10.5281/zenodo.23366156 |
| Clean AI Engineering | The specification formats, the four gates and the cycle that joins the parts abovev0.1.0 | Public | 10.5281/zenodo.23366162 |
Evidence so far
Three agents we didn't build
The airline, retail and telecom agents of τ²-bench, a published benchmark. For each, we wrote a specification from the agent's policy text and tool signatures alone, then judged the same 2,472 runs twice: by τ²-bench's own evaluator and by AgentTwin.
| Domain | Runs | Rule breaks the data shows τ² / AgentTwin | Reads of another customer's record τ² / AgentTwin | False alarms on correct runs |
|---|---|---|---|---|
| Airline | 467 | 188 / 188 | 0 / 50 | 2 of 50 |
| Retail | 873 | 309 / 309 | 0 / 114 | 0 of 114 |
| Telecom the agent's side only | 1,132 | 50 / 491 | 0 / 114 | 0 of 114 |
- The rule breaks visible in the data were caught (except duplicate bookings, a known gap: 2 runs), and each incident names the rule broken, for example
cancel_reservation:EHGLP3:precondition:cancellation_basis. - Privacy reads leave no trace in the final data, so a final-state judge cannot see them. AgentTwin's ownership judge does.
- In telecom, τ²-bench's tasks check the customer's phone, not the carrier's records; AgentTwin checks the records.
- Every disagreement has a stated cause. The run fails if one cannot be explained, so causes cannot be added afterwards.
- The stand-in answered all 8,126 checked tool calls byte-identically to τ²-bench's own environment (1,512 airline, 4,621 retail, 1,993 telecom).
- The method also surfaced faults in the benchmark: a retail task whose only expected action τ²-bench's own code refuses, and a tool that records every changed item at the last item's price.
These runs are each task's correct actions plus deliberate, hand-made mutations of them, with no model calls. They show what each judge can see, not how often a real agent does each thing.
Two reference agents of our own
A real handoff fault
A model-played customer said "my jacket is lost need help". The support agent's lost-parcel rule did not fire, and when the model failed the conversation ended without reaching a person.
Recorded as T-114 →180 of 180 injection cases held
60 from PyRIT, 60 from AgentDojo and 60 built-in, planted in records both agents read, and judged by what changed in the world. Run with a scripted model, so this tests the agents' safeguards, not a model's resistance.
The run record →Coverage gaps, measured
Hand-written test scenarios covered 81 of 468 value pairs and 3 of 16 state transitions (support), and 67 of 403 and 3 of 8 (claims). A property test found a stand-in that accepted any value.
Recorded as T-113 →The lab, with scripted fixers
An honest fixer's change is deployed and the next day is clean. A fixer that weakens its own test, and one that narrows the rule behind the incident, are both refused and handed to a person.
The run reports →- AgentTwin cannot yet express rules about a change (new flights versus old, fewer bags than before), facts nobody records (why a customer is cancelling), or a customer who acts with their own tools (telecom's phone).
- Conversational duties, such as getting an explicit yes before a write, are not state, and no judge here reads them.
- No real-model fixer has completed a fix yet. One free-tier attempt ended on provider limits before writing anything; real fixer runs are month 1 of the study below.
The two studies proposed
Both are designed and the tools exist. The measures will be published before the main runs.
(a) AgentTwin 1.0
Seeking fundingMake it general and validated, so anyone can assess an agent they didn't build.
- A stable, published world format, and worlds built from MCP and OpenAPI tool descriptions.
- Judge validity: agreement with human expert labels and with τ²-bench's judge, with every disagreement explained.
- Assessment reports on three or more public agents, using one protocol.
- A 1.0 release with a DOI, and tutorials.
(b) The fixer study
Seeking fundingThirty simulated days with planted faults. Record exactly what the fixer does.
- Three routes to weaken a check: a test, a monitor, the specification.
- Three conditions: checks writable but recorded; protected; protected plus a reviewer agent.
- Measures: faults truly fixed, weakening tried and landed, protection crossings, honesty mismatches, repeated runs.
- A dataset with a DOI, a short paper, and a protocol any team can run on its own fix loop.
Openness and safety
- Open: code under Apache-2.0, specification text under CC BY 4.0, results and run records in the repositories.
- Sandboxed: agents act only on a simulated world. No real systems, customers or money.
- Synthetic people: customers and their records are generated.
- A person approves every deploy. The lab refuses a fix that weakens a check before anyone is asked.
Reproduce it
Each needs uv and git (and Node, for the catalogs). No model calls and no API keys.
git clone https://github.com/dataagentsai/agenttwin && cd agenttwin uv run --extra dev pytest
cd agenttwin/examples/tau2-airline ./run.sh -m tau2_airline.offline # retail and telecom: same pattern
for r in agenttwin-lab agenttwin reference-agent clean-ai-engineering; do git clone https://github.com/dataagentsai/$r; done cd agenttwin-lab && uv sync --extra dev uv run agenttwin-lab day --world worlds/clothing-day.yaml --repeat 3
for r in ai-assurance-catalog ai-harness-catalog; do # beside the repos above git clone https://github.com/dataagentsai/$r; done (cd ai-harness-catalog && npm ci) cd reference-agent && uv run --extra dev python -m pytest
Who
Basant Choudhary is a principal data and AI engineer in Bengaluru with 23 years in software and data engineering at Oracle, Sabre, Publicis Sapient and Shell. Basant builds production AI agents (insurance claims, cost root-cause analysis, platform operations) and authored the open Agent Harness Framework: two specifications of 118 capabilities and 118 testable obligations, a reusable harness library, AgentTwin, and two reference agents that pass four automated gates.
Contact: basant@cleandataengineering.ai
Last updated 14 Oct 2026.