Execution Backends
The orchestration engine supports three execution backends. The planner and step contracts stay the same; only how each step runs changes.
Comparison
| Backend | Env value | How steps run | Good for |
|---|---|---|---|
| CrewAI in-process | inprocess (default) |
Whole crew in one Python process | Local dev, single machine |
| Subprocess | subprocess + AGENTIC_SUBPROCESS_WORKERS=1 |
One python main.py --execute-step worker per step |
Step isolation, container testing |
| Kubernetes | kubernetes |
K8s Job per step, PVC run store | Production, multi-tenant |
In-process (default)
No extra configuration. CrewAIExecutionBackend runs agents in the coordinator process.
# implicit default
AGENTIC_EXECUTION_BACKEND=inprocess
Subprocess
Each step spawns a worker subprocess that reads a StepSpec JSON file and writes result.json to the shared run store.
export AGENTIC_EXECUTION_BACKEND=subprocess
export AGENTIC_SUBPROCESS_WORKERS=1
export AGENTIC_RUN_STORE_PATH=/tmp/agentic-run-store
python main.py --dynamic "Smoke test subprocess isolation"
Status: F0–F4 complete — subprocess path proven in CI.
Kubernetes
Each step becomes a one-shot K8s Job mounting a ReadWriteMany PVC at AGENTIC_K8S_RUN_STORE_MOUNT (default /run/store).
export AGENTIC_EXECUTION_BACKEND=kubernetes
export AGENTIC_K8S_WORKER_IMAGE=agentic-orchestrator-worker:local
export AGENTIC_K8S_RUN_STORE_PVC=agentic-run-store
export AGENTIC_RUN_STORE_PATH=/run/store # coordinator must share the same store
Deploy manifests: agentic-orchestration-tool/deploy/k8s/.
| Concern | Notes |
|---|---|
| MCP stdio | Worker-native stdio or cluster sidecar gateways |
| Run store | hostPath (kind), NFS, or GKE Filestore |
| CI | kind e2e with stub worker image |
Status: K3 MVP implemented (kind CI e2e). K5 operational polish (warm pool, structured logs) optional.
Run store layout
{AGENTIC_RUN_STORE_PATH}/{run_id}/{step_id}/result.json
{AGENTIC_RUN_STORE_PATH}/{run_id}/{step_id}-spec.json
Workers and coordinator must share this path (bind mount, PVC, or NFS).
What stays the same
- YAML agent and MCP catalogs
- Planner JSON plan format
- Session, KB, learning, and QA hooks
--execute-stepworker contract
See Architecture for the full pipeline.
Deep dives: Dual execution framework (F0–F5 code seam) and Kubernetes execution upgrade (cluster delivery, K3–K5).