Wandb Improve

Self-Improvement Workflow for Promus

You have access to the W&B MCP Server. Follow this loop:

Step 1: Run current evals

Run cd evals && uv run python scripts/run_eval.py to get baseline scores.

Step 2: Analyze results via W&B MCP

Use the W&B MCP tools to:

1. Query the latest evaluation results: "Show evaluation scores for promus/task-agent"
2. Query recent traces: "Show the last 20 task processing traces, sorted by latency"
3. Identify patterns: Which intents fail? Which tools are misrouted? What's slow?

Step 3: Diagnose issues

Read the current prompts in evals/scripts/run_eval.py:

• INTENT_CLASSIFIER_PROMPT — intent classification system prompt
• TASK_PLANNER_PROMPT — task planning system prompt

Compare prompt instructions against the failures found in traces.

Step 4: Improve

Edit the prompts to address identified issues. Common improvements:

• Add few-shot examples for misclassified intents
• Clarify tool selection rules
• Add tone guidance for drafting
• Adjust temperature for consistency

Step 5: Re-evaluate

Run cd evals && uv run python scripts/run_eval.py again to generate new scores.

Step 6: Report

Use W&B MCP create_wandb_report_tool to create a comparison report showing before/after metrics. Document what changed and why in a commit message.