Live collection and branching

Playback, live collection, and branch execution

These three get confused, so it's worth being precise:

• Playback steps through history that's already saved. No model is called.
• Live collection records an agent application while it runs and streams each new event to the explorer.
• Branch execution takes a saved conversation up to some point and continues it with real model and tool calls, using a compatible runtime.

Run the CrewAI example

The integration is tested against CrewAI 1.15.23 on Python 3.12. From the repository root:

python3.12 -m venv .venv-crewai
source .venv-crewai/bin/activate
python -m pip install -r examples/crewai/requirements.lock
python -m pip install --no-deps -e .
export OTEL_SDK_DISABLED=true
export CREWAI_TELEMETRY_DISABLED=true
export CREWAI_TRACING_ENABLED=false
export CREWAI_STORAGE_DIR="$PWD/data/crewai-storage"
swarm-lens --data data --port 8765 \
  --runtime examples.crewai.demo:create_runtime \
  --trace-tools examples.crewai.demo:trace_tools

In a second terminal, with the same environment activated:

python -m examples.crewai.demo --offline --url http://127.0.0.1:8765

--offline runs the real CrewAI loop but swaps in fixed, deterministic responses, so it's free. Drop the flag (and set OPENAI_API_KEY) to use a real model, which you'll pay for.

Instrument your own crew

Wrap your kickoff call with swarm_lens.integrations.crewai.observe, passing the crew, its inputs, and a registered runtime. Your application stays in charge of the crew factory, the tool implementations, and its own revision. The integration example walks through the full setup, including observer configuration and which kinds of state aren't supported yet.

Continue a saved conversation

Select an event, pick a fork action, check the runtime preview, and choose Create and run. Whatever you changed (a goal, a prompt) applies only to the new branch. The parent branch, including everything after the fork point, is left alone.

How the run resumes depends on where the trace came from. For a trace captured natively from CrewAI, the adapter can pick up at a task boundary or restart the current task with the recorded context. An imported trace gets rebuilt as a CrewAI crew: ACIArena LLMDebate keeps its debater-then-aggregator order, and anything else takes turns round-robin. Either way, this is a new run that starts from the recorded messages. The original framework's internal state isn't restored.

Tools need some care. They have to be registered explicitly, and a saved tool result isn't the same as calling the tool again. Forking also won't undo anything a tool did to files or external services, and if a task restarts it may call those tools a second time. The preview tells you, before anything runs, which tools are available and which agent will act next.