The full, reproducible evidence behind the integration page. A real crewai 1.15.18 crew, run in-process with the Mission SDK, measured on delegation authority, a governed action, and a real end-to-end crew decision. Every comparison is classified — this is conformance, not a scorecard.
| Experiment | Native role | ReDevOps addition | Result | Classification |
|---|---|---|---|---|
| CR-A Delegation authority | delegate runs its tools | envelope narrows on handoff | 3/3 widening denied · 0 wrong | REDEVOPS_DELTA |
| CR-A In-scope preserved | — | authorized work admitted | 2/2 admitted | PARITY |
| CR-B Governed action | crew recommends | authority · approval · replay | 0 unauthorized · 0 duplicates | REDEVOPS_DELTA |
| CR-C Real crew (n=6) | 2-agent crew decides | wrapped as a Mission | accuracy 1.000 (preserved) | PARITY |
| CR-C Grounding gate | crew always answers | evidence-grounded auto-action | 6/6 grounded (0 routed here) | EXPECTED_IMPLEMENTATION_DIFFERENCE |
| CR-D No pooling | crew shares tools | per-agent envelope | grants don't sum / borrow | REDEVOPS_DELTA |
| CR-E Telemetry | crew/agent spans | Mission-root nesting | preserved (nested) | REDEVOPS_DELTA |
Real crewai agents form a delegation chain: a Compliance Officer (grants kyc:review, case:note) delegates to a Risk Analyst, who delegates to a Junior Analyst. Each delegate's effective grants are the intersection with its delegator's. When a delegate attempts an action outside that narrowed envelope — freezing an account its delegator never held, releasing funds no one in the chain holds — ReDevOps denies it. Native CrewAI would run all 3/3 of those authority-widening actions; ReDevOps denies all three with 0 wrong, while admitting 2/2 legitimately-scoped actions, including a freeze delegated from a Senior Officer who actually held it.
The crew recommends a remediation; it runs only through a Mission. Missing grant ⇒ 0 unauthorized actions. Grant but no approval ⇒ parked at a human gate, 0 side effects. Approved ⇒ fires exactly once. Replay ⇒ 0 duplicate side effects. A crew consensus is not authority to act.
A real 2-agent crew (a risk analyst hands off to a decision officer, sequential process) triages six AML alerts and returns a decision with a rationale. Accuracy is 1.000 on this set, and ReDevOps wraps the crew as a Mission without changing what it decides — the framework's job stays the framework's. Every auto-actioned decision must be grounded in the alert's decisive signal; here all six are grounded so none are routed. That grounding gate is the same one the LlamaIndex slice shows biting hard (6/6 unsupported answers abstained).
Each run writes a reproducibility record — framework version, model ids, per-finding classification, and a content-addressed result digest.
// reproducible · one Python env · github.com/redevops-io/redevops-integrations/integrations/crewai · integration page · architecture article