Training Crews

Crew training is an interactive human feedback loop: you run the crew, review each agent’s outputs, add corrections or preferences, and repeat across iterations. Those rounds teach the crew—without editing Python—what “good” looks like for your domain, so later runs align more closely with how you want work delivered.

CLI: train and save a model

From a CrewAI project directory:

crewai train -n 5 -f trained_model.pkl

-n sets how many training iterations to run; -f is the output file path for the serialized training data.

Programmatic training

Call train() on the same Crew instance you would kickoff():

crew.train(
    n_iterations=3,
    inputs={"topic": "AI Agents"},
    filename="trained_agents.pkl",
)

Pass inputs the same way you do for kickoff() so tasks with {placeholders} resolve during training runs.

The training loop (step by step)

1. The crew executes normally (tasks run in process order).
2. You see each agent’s output for the current step.
3. You provide human feedback (what to keep, fix, or emphasize).
4. The agent revises its output using that feedback.
5. You repeat until you have completed n_iterations rounds.

Each iteration sharpens behavior through your corrections rather than by changing agent definitions in code.

What gets saved

The training artifact (for example trained_agents_data.pkl) stores a consolidated bundle per agent, including:

• Suggestions distilled from your feedback across iterations
• Quality scores associated with those rounds
• Summaries that capture what the crew should do differently next time

Keep this file in version control or a shared artifact store if you want the same “institutional” behavior everywhere the crew runs.

Automatic application on future runs

When you load or point the crew at trained data, agents automatically append their saved suggestions to task prompts in subsequent executions. You do not manually merge strings: the framework injects the learned guidance so each task benefits from prior human review.

Crew testing (benchmark-style runs)

Testing is separate from training: it runs the crew multiple times and auto-scores outputs so you can compare stability and quality without typing feedback each time.

crewai test -n 5 -m gpt-4o

• -n: number of test executions
• -m: model used for scoring (and execution, per your CLI/project defaults)

Each task is scored on a 1–10 scale; you get aggregate views across runs.

Example testing output (conceptual)

Exact column names depend on the CLI version, but expect per-task scores, averages, and timing so you can spot regressions after prompt or tool changes.

When to use training

• Tune behavior without code churn — nudge tone, structure, and policy through feedback instead of rewriting backstory every time.
• Capture institutional knowledge — encode how your team reviews outputs so new runs match internal standards.
• Pair with testing — after training, use crewai test to see whether scores stay high across repeated executions.

Key takeaways

• Training = interactive loop: run → review → feedback → revise → repeat for n_iterations.
• CLI: crewai train -n 5 -f trained_model.pkl; code: crew.train(..., filename="trained_agents.pkl").
• Saved pickle holds suggestions, scores, and summaries per agent; future runs append that guidance to tasks automatically.
• Testing: crewai test -n 5 -m gpt-4o runs the crew repeatedly with 1–10 task scores, averages, and execution time for regression-style checks.