Build your first bounded experiment

Published and revised: October 13, 2026 · Version 1.0.0

Go from an architecture choice to a small simulated run and a reviewable failure log.

Goal and proposed time

Decide whether an agent is warranted, specify its boundary, and inspect one normal case plus two meaningful failure cases.

Allow 140–180 minutes, including selected reading, making the outputs, and review. This is a proposed study budget, not measured completion time. Split the path across sessions if needed; stop at a checkpoint rather than rushing an evidence check.

Prerequisites and inputs

• No agent framework or coding experience required
• Be able to distinguish a proposed action from an action actually performed

Use only the linked guides and their public or synthetic fixtures. A text editor, calculator, and timer are enough. No provider connection, account change, publication, or outside communication is required.

Work through the path

1. Choose an agent, a workflow, or a simpler solution

Read the guide. Write the outcome and use the five gates. Keep a fixed workflow unless the next-step decision genuinely depends on new evidence.

Carry forward: A one-page architecture choice with the reason to test, or retain, the simpler option.

2. Map an agent's permissions before it acts

Read the guide. Separate reading, draft creation, and external change. Fill the matrix for the workshop-kit simulation; do not add a purchasing capability.

Carry forward: An action-by-action matrix with a named exercise operator.

3. Write a tool contract that can reject a bad request

Read the guide. Review the room-lookup example to learn input, authorization, and result checks. Then apply those distinctions to the three existing kit tools without changing their fixtures.

Carry forward: A validator checklist that preserves the kit exercise’s exact fields and limits.

4. Rehearse a bounded tool-using experiment

Read the guide. Run the normal kit case as a manual simulation. Keep the operator’s actual returns and count every request, including rejected ones.

Carry forward: A saved first attempt and a supported recommendation or precise stop.

5. Evaluate an agent and learn from its failures

Read the guide. Review T01, then run T04 and T07 fresh. Score the visible action path as well as the final answer; keep remaining cases as an explicit next test set.

Carry forward: Three case records, earliest visible divergences, and one bounded repair if needed.

Worked checkpoint

For T01, the expected shortage is three kits and the fictional quote is 1550 cents. T04 must leave shortage and cost unknown because the requirement is missing. T07 preserves the ordinary calculation while ignoring the instruction hidden in the stock note. These are different passing behaviors, not three variations of a successful purchase.

Output

One architecture decision, a permission matrix, a read-only tool contract, and a three-case rehearsal log with an explicit keep/revise/hold decision.

Exit checks

• The simpler baseline is named and compared fairly
• Every operation has an allowed target and an enforcement owner
• Malformed requests and invalid results are rejected
• T01, T04, and T07 have separate observed outcomes; proposed and executed actions are distinguished
• Unknowns and failed runs remain visible

Decide what happens next

Keep the result when the checks pass. Revise a specific failed check in a new draft, preserving the first attempt. Hold any part that needs missing evidence or authority; an honest hold can complete this learning path even though the wider real-world task remains unfinished. Do not call an authored answer key an observed result.

IIEN means infer intent and extrapolate needs here: use it to notice a useful next question and label assumptions. It does not authorize new actions, expand access, or tune model weights.

Sources and scope

This is an editorial sequence of the linked existing guides. Their source sections support the individual methods; the sequence, checkpoints, and study budget are original teaching choices. This path does not claim certification, deployed capability, or measured performance improvement.