Training And Evaluation

Operator calls crew.train(n_iterations=1, filename='trained.pkl', inputs=...) and expects meaningful improvement.

n_iterations=1 is rarely enough — training is the iterative refinement of agent prompts via human feedback. Budget 5-20 iterations on a representative input set; verify per-iteration the human feedback is actually reaching the prompt. Do not treat training as a one-shot.

Training saves output to trained.pkl. The operator wants to inspect what actually changed.

trained.pkl contains per-agent / per-iteration prompt refinements. Diff against the baseline prompt to surface what training learned. Treat the artifact as code: review in PR, version, and gate behind tests. Don't ship a trained artifact whose changes you can't explain.

During crew.train(), CrewAI pauses and prompts the operator for feedback after each task in each iteration.

Human feedback is the training signal — provide specific, actionable corrections tied to expected_output. Generic 'good' / 'bad' is wasted signal. Capture the feedback in a structured store for replay and review. Verify feedback is applied by inspecting the saved prompt artifact.