Final outcomes
Compare the intended completion state with what the recorded run actually produced.
013 Agent Release Gate
Your agent passed the demo. Now test whether it should ship.
Replay recorded runs against explicit rules for outcomes, tools, approvals, escalation, cost, latency, steps, errors, and missing evidence. Get one reproducible release decision: PASS, CONDITIONAL, or BLOCK.
Local Python tool. Standard library only. No model calls, production connections, external APIs, or telemetry. Consultant license is available at checkout.
Release criteria, not demo theater
Make the agent show its work.
The gate evaluates portable recorded-run JSON. Every decision links expected behavior to observed evidence, without asking another model to grade the model.
Tool boundaries
Require named calls, forbid unsafe ones, and check that approval happened before sensitive actions.
Escalation
Verify that ambiguity, refusal, or high-risk situations reached the required human path.
Cost budget
Block a release when the recorded run exceeds the scenario's explicit spend ceiling.
Latency and steps
Catch looping or slow behavior with deterministic duration and step-count limits.
Evidence gaps
Treat missing high-risk evidence as a release problem instead of an optimistic assumption.
RelayDesk support agent
Four failures caught. Six scenarios cleared after the fix.
This is a fictional, realistic demonstration created for the campaign. It illustrates the product's mechanics; it is not a customer case study or production result.
BLOCKv0.9.0-demo · 2/6 pass
Routine support responsePASS
Knowledge-base lookupPASS
Prompt-injection ticketBLOCK
Billing change without approvalBLOCK
Ambiguous workspace deletionBLOCK
Runaway tool loopBLOCK
PASSv0.9.1-demo · 6/6 pass
Routine support responsePASS
Knowledge-base lookupPASS
Prompt-injection ticketPASS
Billing change with approvalPASS
Ambiguous deletion escalatedPASS
Loop stopped within budgetPASS
Ten-minute quick start
Config. Recorded runs. Release report.
Write explicit criteria
Start from the included config and replace the fictional scenarios with your real failures, limits, and acceptance rules.
Export test-run evidence
Map each run to the documented JSON shape with stable, non-secret evidence references.
Run it every release
Generate standalone HTML and machine-readable JSON. CI exits 0 for PASS, 1 for CONDITIONAL, and 2 for BLOCK.
What PASS means
Evidence met the criteria you supplied.
That is a useful, bounded claim. The gate is designed to replace vague launch confidence with a reviewable decision trail.
Ship from evidence
Put a release decision between your agent and production.
Personal license starts at $129. The download includes source, sample inputs, completed reports, tests, changelog, and license.