Outcome DrivenRead the full guide →

EVIDENCE BEFORE EXPANSION / FREE POCKET PRODUCT

Agent Workflow Evidence Kit

Measure accepted outcomes, expose missing data and make a defensible pilot decision.

Start the worksheet ↓Download the guide (PDF)Get the complete kit (ZIP)
Original editorial illustration for Agent Workflow Evidence Kit

A small kit.
A useful finished result.

Design a small paired experiment for an MCP or WebMCP workflow. Choose the decision and scoring rule before the run. The included dataset is fictional and intentionally imperfect; it demonstrates how to report a missing duration without manufacturing a time-saving claim.

  1. Agree on the decision

    Complete the hypothesis sheet before seeing results. Define a reviewed task, a realistic baseline and a primary acceptance condition.

  2. Prepare paired cases

    Use the same underlying tasks for both conditions. Freeze or record source versions, vary execution order and define the deadline.

  3. Capture the full run

    Keep requests, tool results, artifacts, corrections and review notes. Record human interventions instead of crediting them to the agent.

  4. Score and calculate

    Apply the rubric to every eligible case. Keep missing values null and label the denominator and subset used for each calculation.

  5. Write a bounded conclusion

    Use the memo to state what the evidence supports, its limitations and the next experiment. Retain failed runs when making improvements.

Agent Workflow Evidence Kit: step-by-step workflow
Original instructional graphic. Open it to read at full size.

YOUR WORKING CANVAS

Make it specific.

Write your own notes, or load the fictional example to see a completed version. Download your work before closing this page.

Your notes stay in this page's memory. Nothing is submitted, saved to an account or sent to an AI service. Closing or reloading the page clears your work.

Name a specific pilot or product decision, not a vague goal to prove AI works.

Include the user-visible artifact and required human review.

Record sources, model, prompt, tool version and other differences.

Use an inspectable pass condition. Critical errors should not be offset by formatting points.

Include ordinary work and cases likely to reveal weaknesses.

Use the same boundaries and include review, corrections and interventions.

Keep unknown values missing. State which denominator each metric uses.

State the evidence needed, owner, limitations and next review date.

Your export preview

EDITABLE COMPANIONS

Take the next step.

KEEP THE WHOLE KIT

Work at your pace.

The ZIP includes the PDF guide, the original graphic, this offline worksheet and all editable templates. Open index.html in a browser after extracting it.

Use the materials for your own project or client work. Check and adapt the examples; they are fictional demonstrations.

Download the complete kit

Return to the article →