Robochoke · Agent-readiness report
https://example.com/checkout
Agent goal: Buy the Denim Jeans for under €80 without paying
Completed up to payment — the run halted at the configured safety boundary.
Outcome base 85 − 0 in penalties · Formula v1
Friction findings
2 findings — ranked, most costly first.
Named issues the agent hit, ranked. Each names the exact element, suggests a code-level fix, and links to the moment it happened in the replay below.
- Unstable click target — 'Add to cart' buttonstep 4
The agent's first click on the 'Add to cart' button missed during a layout shift — the miss is recorded in the step's error log — and it retried on the settled page before succeeding.
ElementSuggested fixReserve the container's dimensions (fixed aspect-ratio or min-height) so the button cannot move while the page's assets finish loading.
- Slow interaction on the 'Full name' fieldstep 5
The agent dwelled for 16.2s on the contact name field, deliberating over whether any input was needed to satisfy the goal's under-€80 constraint without paying.
ElementSuggested fixMark the field optional inline (or pre-fill it) so an agent does not pause to reason about whether completing it is required.
Run replay
Every step the agent took, with its reasoning. Select a step to inspect it, or jump to the recorded video for ground truth.
Sensitive typed values are masked before data leaves the server — the real values are never sent to your browser.
Score breakdown
score = max(0, outcome_base(verdict) − Σ penalties)
No friction penalties — clean run.
Formula version: v1 · Deterministic — recomputed from stored signals when the formula changes.
Why this verdict
The agent filled the contact details and reached the order summary with the €64 total under the €80 constraint, then the safety gate halted the run before payment.
Verifier confidence: high
Verdict constrained by harness: gate halt.
Run details
- Run ID
- seed-replay-full
- Target URL
- https://example.com/checkout
- Boundary preset
- halt-before-paymentThe agent was required to stop before money moved.
- Agent model
- claude-sonnet-4-6
- Verifier model
- —
- Started
- 2026-07-22T00:00:00Z
- Duration
- 3m 12s
- Steps used
- 6 / 50
- Signals
- Backtracks 0 · Retries 0 · Hesitations 0 — threshold 15s
- Formula version
- v1
Artifacts
How this score works
The agent-readiness score is deterministic: the same run data always produces the same score. It starts from an outcome base set by the run's verdict — completed 100 · stopped at boundary 85 · blocked by bot defense 50 · incomplete 40 · failed 30 — then subtracts fixed penalties for three friction signals: backtracks (4 points each), retries (3 points each), and hesitations (2 points each). The score is floored at 0.
What it measures: how easily a goal-driven AI agent could work through this flow. What it doesn't: human usability, visual design, performance, SEO, or accessibility.
This report reflects a single run — treat it as a strong signal, not a statistical guarantee. When the formula version changes, all historical reports are recomputed so scores stay comparable.
Score rubric
Outcome base by verdict
| Verdict | Base points |
|---|---|
| Completed | 100 |
| Stopped at boundary | 85 |
| Blocked by bot defense | 50 |
| Incomplete — budget exhausted | 40 |
| Failed | 30 |
Friction penalties
| Friction signal | Points each |
|---|---|
| Backtracks | 4 |
| Retries | 3 |
| Hesitations | 2 |
Single-run variance
Agent runs are stochastic — the same site can produce a different score on a consecutive run, and run-to-run variance has not been quantified yet. Treat this score as a strong single-run signal, not a statistical guarantee; the findings above are this report's durable, evidence-bound output. Multi-run audits (planned) will separate site friction from agent flakiness statistically.