← all benchmarksIntegration page →
Benchmark · integrations · CrewAI

CrewAI × ReDevOps — evidence

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.

Frozen environment
framework crewai 1.15.18 — real Agents / Crew / BaseTool, sequential & hierarchical process runtime Mission SDK 0.2.0a0 · Context Runtime 0.2.0 — one Python 3.12 env, in-process answer model gpt-4o-mini (crew synthesis, CR-C only) — identical across arms scenario AML / KYC compliance crew (delegation chains + a governed remediation) authority per-agent grant envelopes; a delegated agent's effective grants = intersection of the chain scoring authority-widening actions denied · exactly-once side-effect witness · crew decision accuracy

Results — CR-A through CR-E

ExperimentNative roleReDevOps additionResultClassification
CR-A Delegation authoritydelegate runs its toolsenvelope narrows on handoff3/3 widening denied · 0 wrongREDEVOPS_DELTA
CR-A In-scope preserved—authorized work admitted2/2 admittedPARITY
CR-B Governed actioncrew recommendsauthority · approval · replay0 unauthorized · 0 duplicatesREDEVOPS_DELTA
CR-C Real crew (n=6)2-agent crew decideswrapped as a Missionaccuracy 1.000 (preserved)PARITY
CR-C Grounding gatecrew always answersevidence-grounded auto-action6/6 grounded (0 routed here)EXPECTED_IMPLEMENTATION_DIFFERENCE
CR-D No poolingcrew shares toolsper-agent envelopegrants don't sum / borrowREDEVOPS_DELTA
CR-E Telemetrycrew/agent spansMission-root nestingpreserved (nested)REDEVOPS_DELTA
CR-A · a delegate cannot widen authority (the flagship, no LLM)

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.

CR-B · governed action

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.

CR-C · a real crew, end-to-end

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).

Reproduce

git clone https://github.com/redevops-io/redevops-integrations cd redevops-integrations/integrations/crewai ../../.venv/bin/python conformance.py # CR-A..E + classification tally ../../.venv/bin/python examples/delegation_authority.py # CR-A (no LLM; real crewai agents) ../../.venv/bin/python examples/governed_delegation.py # CR-B (crew recommends, Mission governs) ../../.venv/bin/python examples/crew_decision.py # CR-C (real 2-agent crew, gpt-4o-mini, cached)

Acceptance bundles

Each run writes a reproducibility record — framework version, model ids, per-finding classification, and a content-addressed result digest.

results/cr_a_delegation_authority.json crewai 1.15.18 · no LLM result_digest sha256:6a89200be73f2d63 results/cr_b_governed_delegation.json crewai 1.15.18 result_digest sha256:ef142fad45ba87e1 results/cr_c_crew_decision.json crewai 1.15.18 · gpt-4o-mini result_digest sha256:1c6d8216d26650c5 adapter_commit 16060d4 · mission-sdk 0.2.0a0 · context-runtime 0.2.0
// conformance summary   CrewAI suite over all findings: 7 REDEVOPS_DELTA · 2 PARITY · 1 EXPECTED_IMPLEMENTATION_DIFFERENCE · 0 BUG. The PARITYs are in-scope delegation preserved and the crew's own decision accuracy (wrapped, unchanged); the EXPECTED difference is the grounding gate that routes nothing on this clean set. Neither is a win nor a bug.
Limitations. The authority scenario is a constructed AML/KYC fixture (deterministic, no LLM) that isolates the delegation-narrowing mechanism, not a live compliance system. CR-C's crew runs on a small, clean six-alert set, so its grounding gate routes nothing here — the gate's value shows when a fluent agent decides without support. Reported straight; nothing averaged into a single headline.

// reproducible · one Python env · github.com/redevops-io/redevops-integrations/integrations/crewai  ·  integration page  ·  architecture article

CrewAI is a trademark of crewAI, Inc. Independent integration, not a CrewAI product or endorsement.