agentverity 0.8.6

Decision stability and coverage checks for AI agent tests

AgentVerity

Decision stability and coverage checks for AI agent tests.

AgentVerity qualifies tests for model-backed components that choose from a finite, reviewed set of decisions. Examples include routers, approval or policy gates, and supervisors that select the next agent or tool. It compares the named decision, exposed as verdict, rather than harmless changes in explanation text.

Use another evaluator for open-ended chat with no reviewed decision or ordered tool-path contract. AgentVerity is alpha, with documented pre-1.0 guarantees.

Read the design story: Introducing AgentVerity: What Does a Green Agent Test Prove?

Is it a fit?

AgentVerity fits when all three conditions hold:

• each run exposes a named decision or a reviewed tool or handoff path
• repeated trials can start from equivalent state in isolated sessions
• you can supply deliberately varied test inputs that should reach different valid decisions

Good targets include support and payment routers, fraud triage, approval and policy gates, incident routing, multi-agent supervisors, and bounded tool selectors. In a larger agent, test the step that owns the decision or the final pipeline decision. Test both when each is a release contract.

It is not an end-to-end quality evaluator for chat, RAG answers, generated content, coding agents, or research agents. If one of those systems also emits a reviewed route, approval, escalation, or tool path, AgentVerity can qualify that decision layer, not the open-ended work around it.

See the applicability checklist and exact limits.

Try it

pip install agentverity

from agentverity import from_callable, run

def route(ticket: str) -> dict[str, str]:
    # Deliberate defect: every ticket takes the same route.
    return {"text": "route: general", "verdict": "general"}

agent = from_callable(route)
result = run(agent, inputs=[
    "my card was charged twice",
    "the app crashes on login",
    "where is my refund",
    "the checkout button is the wrong colour",
])

print(result.headline)

NOT TRUSTWORTHY - the agent answered 'general' on 100% of the probes,
so a pass says more about the probe set than about the agent.

Those deliberately varied test inputs form the probe set. AgentVerity asks:

• Decision stability: does one case reach the same decision across isolated reruns?
• Decision coverage: do the cases reach more than one decision?

A green quality score answers a different question: whether the selected answers were right. AgentVerity checks how much evidence that score rests on. It complements DeepEval, promptfoo, AgentCore Evaluations, and ordinary assertions rather than replacing them.

Together, the checks guard against two failure modes:

• Vacuous green: every supplied assertion passes, but the test inputs reach only one decision.
• Regression trap: that narrow run becomes the baseline, so later changes to untested decisions remain invisible.

AgentVerity qualifies the evidence from this run. It does not certify the agent as correct, safe, or fully covered.

Why rerun counts are harder than they look

Three or five reruns chosen by convention can support the wrong conclusion. In one deterministic example, 36 non-overlapping rerun comparisons produced no decision changes. That was still too little evidence to certify a change rate below 5%. The honest result was undecided, not unstable. Certifying that threshold with no observed changes needed 73 comparisons.

The decision rule is:

• pair reruns without reusing an output
• calculate a 95% Wilson interval around the observed decision-change rate
• report deterministic when the upper bound is below the tolerated rate
• report stochastic when the lower bound is above the tolerated rate
• report undecided when the interval spans that tolerance

Wilson intervals are established statistics. AgentVerity's design choice is to use one as a three-outcome release rule rather than force an underpowered run into stable or unstable. Non-overlapping pairs avoid making the sample look larger than the target calls justify, while the requested tolerance determines the call budget before execution.

The default balanced precision calculates that budget automatically.

See the executable helper, arithmetic, and exact API mapping.

The evidence gate

A baseline is a reviewed set of expected decisions for later versions. AgentVerity refuses to save one until calls complete, decisions are stable enough, the probe set crosses a decision boundary, and a person approves the reference outputs.

The bundled payment-dispute example runs two test sets:

python examples/payment_dispute_gate.py

Probe set	Exact-match	Verdict stability	Probe coverage	Baseline
Narrow, 6 duplicate-charge cases	✅ 6/6	✅ verdict-deterministic	❌ blind, 1 route	❌ REFUSED
Repaired, 6 dispute categories	✅ 6/6	✅ verdict-deterministic	✅ 6 routes	✅ ADMITTED

Both score 6/6. The narrow set correctly tests one route, but it cannot justify a system-wide baseline. The repaired set reaches all six routes and can be saved as a versioned snapshot.

Create one through the CLI:

agentverity snapshot \
  --agent mymod:build_agent \
  --inputs seeds.txt \
  --output baseline.json \
  --accept-reference

The same checks run before agentverity check reports differences as regressions. Snapshot files retain SHA-256 input fingerprints rather than raw prompts.

Where it fits

AgentVerity is an evaluation runner, not serving-path middleware. It makes controlled calls with reviewed test inputs:

reviewed cases ---> agent ---> quality evaluator: "Was it right?"
             +---> agent ---> AgentVerity: "Can I trust this test?"
                                      |
                           snapshot or release decision

Use it while developing, on a pull request, before release, or as a scheduled synthetic canary. Do not repeat live customer requests. Results can leave as text, JSON, JUnit XML, or one privacy-minimised OpenTelemetry span.

See CI, telemetry, lifecycle, and multi-agent integration.

Measured AgentCore canary

The optional production example combines a Strands payment router on Amazon Bedrock, DeepEval route-quality checks, AgentVerity, AgentCore Runtime, and CloudWatch.

The London canary recorded 6/6 correct routes, no changes across 36 repeat pairs, all six routes reached, and 78 successful cloud calls with no errors or throttles. Its first run was stable but only 5/6 correct, so the example now requires both quality and evidence before snapshot admission.

This is deployment proof, not an AWS requirement. The zero-dependency callable works with any stack.

Run the production example · Read the measured result

Scope

A trustworthy AgentVerity result means that, for the supplied probe set and chosen tolerance:

• repeated isolated calls provided enough evidence about decision stability
• observed decisions did not collapse onto one highly dominant route
• execution completed and any requested relation checks were meaningful

That is a minimum dynamic adequacy check, not exhaustive branch coverage. AgentVerity does not establish that every declared decision, rare boundary, or high-risk case was tested. Pair it with a declared route inventory, labelled correctness cases, and stricter separate runs for critical decisions.

The next adequacy extension under consideration is an optional declared decision contract. It would report required, observed, and missing decisions separately from the current skew warning. It is not part of the current API.

It also does not judge answer correctness, prove safety, store traces, host a dashboard, or monitor production traffic. Static tools remain useful for declared branches, route schemas, and expected labels. AgentVerity measures the decisions a model-backed or black-box target actually returns.

Documentation

• Which agents fit, and what the result does not prove
• Why arbitrary rerun counts fail
• Integrations and AgentCore validation
• API guide
• Design decisions
• ADR: compare named decisions, not generated text
• API stability and path to 1.0
• Security and data handling
• Contributing
• Release process

Development

pip install -e ".[dev]"
python -m pytest -q
ruff check .

CI covers Python 3.10 through 3.14, lint, package construction, and the generated README evidence. A coverage job enforces at least 90% statement coverage, and the branch-protection CI gate requires that job to pass.

Status and licence

Alpha. Pin a minor series for production use, for example agentverity~=0.8.0. Patch releases preserve the public API.

Apache-2.0. Contributions are welcome through the pull-request workflow.