ReDevOps Integrations

Reference implementations of the ReDevOps Mission Runtime wrapped around the agent frameworks and clouds you already use. Each integration keeps the framework's native agent loop and adds the production runtime properties around it — closure-aware context, authority, approval, replay, verification and governance — then classifies every native-vs-native+ReDevOps comparison. Conformance, not a scorecard.

Live write-ups, architecture diagrams and benchmark tables: redevops.io/integrations. The Mission SDK these build on: github.com/redevops-io/mission-sdk.

Every slice runs a real framework agent, locally — the four hyperscaler slices need no cloud credentials (the SDKs point at any OpenAI-compatible endpoint; cloud identity is modelled as policy). One provider-neutral adapter contract (integrations/common/) is shared by all eight.

The eight integrations

Slice	Framework / cloud	What ReDevOps adds (the boundary tested)	Headline
nvidia	NVIDIA NeMo Agent Toolkit	closure-aware context + governed Mission	SciFact 0.800→0.867 · GDPR 0.412→0.913 · 0 replay dups
pydanticai	PydanticAI	verification/governance beyond the typed boundary	4 schema-valid outputs refused · 0 replay dups
llamaindex	LlamaIndex	the retriever as one representation, composed with closure	code closure 0.320→0.693 · 6/6 unsupported answers abstained
crewai	CrewAI	authority that narrows across delegation	3/3 authority-widening actions denied · no pooling
azure	Azure Semantic Kernel	compliance gate + Entra-composing authority (deny-wins)	3/3 non-compliant deploys denied · 0 replay dups
google	Google ADK / A2A	chain-wide authority + provenance across a multi-hop mesh	2/2 transitive escalations denied · 2/2 unprovenanced excluded
aws	AWS Strands Agents	Mission authority composed with IAM (deny-wins) + shadow	4/4 policy-denied actions denied · 500 households, 0 side effects
digitalocean	DigitalOcean GenAI	one shared runtime, tenant-isolated	3/3 cross-tenant accesses denied · machinery 5 once vs 15

Each slice's README.md has its own thesis, run commands, expected output and honesty line. Each writes content-addressed acceptance bundles to results/ (framework version, model ids, per-finding classification, result digest) — these are committed, so the measured evidence is public even where a run needs extra setup.

Result classification

Every comparison is one of: PARITY · FRAMEWORK_NATIVE_ADVANTAGE · REDEVOPS_DELTA · EXPECTED_IMPLEMENTATION_DIFFERENCE · BUG (integrations/common/classification.py). The point is a faithful map of where the runtime adds value, not a leaderboard.

Setup

python -m venv .venv && . .venv/bin/activate
pip install -r requirements.txt          # Mission SDK + the framework SDKs you want

The Mission SDK (redevops-mission) is the one dependency every slice needs. Install the framework SDKs for the slices you want to run (see requirements.txt). Model calls go to any OpenAI-compatible endpoint via OPENAI_API_KEY / OPENAI_BASE_URL.

Azure runs in a dedicated venv — Semantic Kernel pins pydantic/openai below the floor the other slices share:

uv venv integrations/azure/.venv --python 3.12
VIRTUAL_ENV=integrations/azure/.venv uv pip install semantic-kernel \
  "redevops-mission @ git+https://github.com/redevops-io/mission-sdk"

Run

cd integrations/<slice>
python conformance.py            # runs the slice's experiments + prints the classification tally
# (azure: use .venv/bin/python)

Reproduction notes (honest)

• Fully standalone (no data, no external repo): all of aws · azure · crewai · digitalocean · google, plus the governed / typed / replay examples of nvidia, pydanticai and llamaindex. git clone, install, run.
• Need a public dataset fetch: nvidia scifact_closure / gdpr_structural and pydanticai closure_verdict read SciFact / GDPR — see workloads/thirddomain/data/README.md.
• Need an external code repo: llamaindex code_closure (LI-A) indexes a target repository via --repo (the measured run used the pinned agentic-os tree); point it at any Python repo to reproduce the mechanism.
• Need a long-context corpus: llamaindex longcontext_answer (LI-B) reads a LiveRAG-style haystack via a benchmark harness not included here; the committed results/li_b_*.json bundle captures the measured run.

In every case the code and the frozen acceptance bundles are public here; only a few runs need the extra input.

License

Apache-2.0 (LICENSE). Framework and cloud names are trademarks of their respective owners; these are independent integrations, not products or endorsements, and use no cloud services.