Mission · local progress
8 min read
Updated
Audit a web app with an agent and verify evidence
Follow a read-only web audit mission: bound the agent’s tools, observe the user journey and verify each claim against evidence.
Direct answer
A bounded audit report ties every finding to a reproducible observation, separates local and public evidence, and lists unknowns. No fix or deployment is implied.
Your journey, step by step
A concept explains, a comparison helps you choose, a recipe composes a screen. This mission connects screens and decisions to the result you must verify.
Have authorization to inspect a named web app, an agent with known tool permissions, and a place outside SkillCodex to keep redacted evidence.
No account required. Only completed step IDs stay in this browser, shared across EN and FR. No external action is executed. Checks are your declarations, not evidence collected by SkillCodex.
Loading progress…
0 / 4 steps declared verified
Resume the next step: Name the audit and its limits1 / 4 · Information
Name the audit and its limits
Screen: the audit brief. Choose a route and user task, such as reopening a saved project. State allowed reads, environment, expected result and forbidden writes. Reuse the safe-agent guide and its existing Pack.
Expected result: The brief names a verifiable result and a read-only boundary.
Common mistake: A broad “audit everything” request cannot define complete coverage.
2 / 4 · Decision
Check the actual tool permissions
Compare the requested read-only scope with the agent’s actual capabilities. Remove unnecessary write access or use an isolated test setup. If authorization or access is unclear, pause.
Expected result: Authorized scope and effective permissions agree.
Common mistake: Instructions alone are not a technical permission boundary.
3 / 4 · External action
Run the authorized observation
In your own agent tool, read the static Pack before copying it unchanged. Provide scope separately. Observe the named journey and capture route, build identity, steps, expected and actual result, with redacted evidence. Do not publish, modify accounts or fix code as part of the audit.
Perform this yourself in your tool, only if authorized. This button updates local progress; it runs no operation.
Expected result: Each finding has reproducible steps and a bounded observation.
Common mistake: A confident agent statement without a trace is not a finding.
4 / 4 · Verification
Replay and classify the evidence
Screen: audit report beside the tested app. Replay the finding. Separate local tests, Git revision, CI, deployment, HTTP and observed user behavior. A healthy endpoint does not prove the task works. Mark unavailable or conflicting proof unknown.
Expected result: The conclusion follows the evidence actually inspected, with gaps named.
Common mistake: Do not turn synthetic observation into human research or device qualification.
Result still to verify
A bounded audit report ties every finding to a reproducible observation, separates local and public evidence, and lists unknowns. No fix or deployment is implied.
Keep evidence in your own environment. SkillCodex certifies neither a real action, human research, nor a mobile device.
01
What you verify
Have authorization to inspect a named web app, an agent with known tool permissions, and a place outside SkillCodex to keep redacted evidence.
Keep this
Final checklist
- 01The brief names a verifiable result and a read-only boundary.
- 02Authorized scope and effective permissions agree.
- 03Each finding has reproducible steps and a bounded observation.
- 04The conclusion follows the evidence actually inspected, with gaps named.
Primary sources
Technical claims in this guide connect to first-party specifications and documentation.
- A practical guide to building agents (opens in a new tab)OpenAI · Operational frame: goal, tools, instructions, and guardrails before granting autonomy.
- Agents and the agent loop (opens in a new tab)OpenAI Agents SDK · The loop stops on output, a tool, a handoff, or a turn limit — which is why the mission must be bounded.
- Trustworthy agents in practice (opens in a new tab)Anthropic · Human control, transparency, and evidence rather than a confidence the agent declares.