Use cases
How Coco governs an agent in one domain at a time, what each page shows, and how to read its numbers.
See how Coco governs an agent in one domain at a time.
Each page takes one kind of agent work and shows the rules, the verdicts, the record, and the measured results. Every page says plainly what's built, what's partial, and what hasn't been measured.
| Use case | What it covers | Where it runs today | What has been measured |
|---|---|---|---|
| Agent payments | Supplier payments, PayTo, and agent wallets. | The demo payment agent on the SDK middleware, in process. | The most measurement: the payment corpus, AgentDojo banking, and the payment failure cases. |
| KYC onboarding | An agent working customer onboarding cases. | The Claude Code hook. | No KYC figure. A functional demo of 11 scenarios. |
| KYB onboarding | An agent onboarding business customers. | No KYB agent runs yet. The draft contracts load in the SDK engine. | No KYB measurement. |
| AI drug discovery | An agent drafting regulatory answers and lab tool submissions. | A demonstration. The lab contracts are unused drafts. | No benchmark. A regulatory demonstration on two questions. |
What a use-case page contains
- The problem: The problem in the domain, and why the checks already in place miss it.
- Control: Rules from the organisation's own policy decide first, and an optional classifier only tightens a verdict.
- Intervention: Each call ends in AUTO_EXECUTE (run), OBSERVE (run and flag), ESCALATE (pause for a person), or DENY (deny with the reason).
- Evidence: Every decision writes a hash-chained receipt, and each compiled rule cites its document, page, section, and clause.
- Policy from the organisation's own documents: How the contracts come from the documents.
- How Coco fits the workflow: Which tool calls the agent makes, and whether the case runs today on the SDK middleware, the HTTP gate, or the Claude Code hook.
- What the guardrails stop: Measured numbers only, each with the command that reproduces it.
- What Coco can't promise yet: The gaps, stated plainly.
How to read the numbers
Each figure names its benchmark and what it counts. A legitimate call held for a person counts as a cost. The benchmarks measure different things:
- Payment transfer corpus: 73,356 payment transfer calls built from 4,371 human-written attacks on a simulated bank teller agent. The rules and its labels both come from the corpus's own policy levels, so it measures how faithfully a policy is enforced, not attack detection.
- AgentDojo: A public benchmark of 588 attacks planted in emails, files, web pages, and payment notes, across four suites, with 97 legitimate tasks. It measures attacks on the agent through what it reads. Rules alone prevented 467 of 588 attacks (run 20260929-185238). Rules plus the classifier prevented 582 of 588, with 18 of 97 legitimate tasks held for a person (run 20260929-184422). The classifier's question in this run was rewritten after seeing the first run's results, so it's tuned to this benchmark.
- tau3 banking: A public benchmark of banking support work. Coco scores each tool call's verdict, not whether the task completes, and the contracts and the violation probes were written by the Coco team.
- Coco's own cases: Scenarios written with their contracts. They show the engine enforces those rules on those records, not whether a customer's records look the same.
Keep in mind: None of these is a production measurement. No figure on these pages comes from a customer's live traffic.
To reproduce the AgentDojo figures from a checkout of the SDK, run:
uv run --no-project --python 3.12 --with agentdojo==0.1.35 \
python bench/agentdojo/export.py --out bench/agentdojo/cases.jsonl
coco eval agentdojo --arms none,rules
coco eval agentdojo --arms judge,hybrid --question v2 --classifier <classifier-url> --workers 16The classifier arms need a running coco-classifier, and vary by a few calls between runs.
Try it
To try Coco on your own policy and records, contact us at trustcoco.ai.
Related Resources
HTTP API
Look up the hosted gate and document service, with authentication, routes, request and response examples, and error codes.
AI drug discovery
How Coco governs an agent drafting regulatory answers and lab tool submissions in drug discovery and product development, and what we have not yet measured.