Evaluate Talon on one workflow
Use this checklist for an initial engineering review. A successful evaluation proves a specific, supported interception point, not universal control over an agent.
1. Record your scope
| Question | Record |
|---|---|
| Which business workflow? | Support reply, coding release, aggregate analytics, contract review, or another single case |
| What could go wrong? | Forbidden action, sensitive egress, unexpected spend or policy-unsafe fallback |
| Which Talon interception point? | Model gateway, MCP proxy, HTTP Action Gateway or delegated runtime |
| Which requests bypass Talon? | Local tool execution, direct API requests, unproxied provider traffic |
| What is the expected business effect? | One allowed mock call, zero forbidden calls, a denied provider request, etc. |
See choose your integration path and control-plane boundaries. Do not mix exact HTTP approval semantics with MCP denial. They are different capabilities.
2. Choose the smallest reproducible proof
- No provider key: 60-second model demo or HTTP action authorization proof.
- Multiple AI use cases: full product demo.
- Coding tool permissions: Copilot CLI integration, which defines its own prerequisites and synthetic release service.
- Existing app: add Talon to the app, beginning with the smallest valid policy.
Match the guide to the version you are actually running. main and the latest release can differ.
3. Run positive and negative tests
- Allowed path: verify the governed request or mock action reaches its expected handler, and capture the policy decision.
- Denied path: verify the forbidden provider request or intercepted tool/action call does not reach its upstream handler.
- Bypass path: run or reason about an unproxied/local action and clearly record that Talon cannot attest to it.
- Evidence integrity: inspect the signed record and verify its integrity. Cross-check it with application logs or the synthetic downstream service counter.
Useful CLI commands for a configured Talon installation:
talon audit list --limit 10
talon audit show <evidence-id>
talon audit verify <evidence-id>
When you have an asserted session, the session guide and coding-agent guide show session-oriented audit and cost commands. A session label alone does not establish an independent authenticated agent.
4. Capture a small evidence packet
| Field | Example of what to write |
|---|---|
| Talon version or commit | Exact deployed binary or source revision |
| Workflow + principal | One authenticated Talon agent/use case; tenant and asserted session when relevant |
| Enforcement boundary | Model API, MCP tools/call, or exact HTTP operation |
| Expected policy decision | Allow / deny / approval required, with the policy rule |
| Actual dispatch result | Zero downstream calls, one mock call, or unknown if not observed |
| Evidence | Record ID, signature verification result and relevant policy snapshot |
| Missing coverage | Unproxied calls, local execution, downstream state not observed |
| Next step | Pilot, integration repair, or explicit no-go decision |
For exported evidence see evidence store, integrity demo and compliance export.
5. Decision
Proceed to a scoped pilot only if the negative test actually prevents the relevant upstream effect, allowed work still succeeds, an operator can explain and verify the decision, and all bypass paths have independent controls or are documented risks.
Talon's HMAC evidence is tamper-evident, not immutable. An approved HTTP action is not automatically dispatched. Runtime cost estimates can differ from final provider charges. Product evidence is not legal or regulatory certification.