Working with tools

An agent uses client tools to read files, edit functions, and run tests. The client checks access and returns results that inform the next step. Agent tool use and feedback.

For this exercise: trace the source, inspect the invented chart, edit the layout, and check the result. Ask the agent to explain edits and report test outcomes.

Choose the tool for the question

Task	Start with	Check afterward
Locate the plotting function	REPO_MAP.md, then rg	Open plot_summary in your language’s source file.
Understand the fixed values	Source constants and the plotting contract	Check assertions for categories, groups, counts, and order.
Diagnose overlap and clipping	The baseline PNG, plotting code, and figure specifications	Locate bar positions, margins, size, and label settings.
Change the layout	plotting/plot_summary.py or plotting/plot_summary.R	Render a new PNG and run that language’s tests.
Review the change	Task brief, source diff, tests, and before/after images	Check the figure specifications and unchanged values.
Review accessibility	Saved image, figure specifications, and summary.alt.txt	Check contrast, grayscale readability, group labels, and a useful text alternative.

Read related source files together. Test the version you intend to keep. With multiple agents, agree on file ownership and which version the checks cover.

Read the result before taking the next step

If a fixed count changes, inspect the assertion and function; keep the expected value. An import failure calls for an environment fix before a layout review.

Check each command’s directory, interpreter, exit status, and result. Open the image rendered from the current source. If the client cannot view it, have a person perform the visual check.

Retry a failed command after addressing its cause.

Check access where the action happens

A file reader, shell, browser, and connector can have different permissions. Client controls enforce access; planning instructions guide behavior. See the setup pages.

Read permissions, sandboxing, and independent work before granting access. Practise declining a request you cannot explain. A sandbox limits covered actions, while meaningful approvals and independent review remain necessary. Keep some agent-free coding practice so that convenience does not become a requirement for doing your work.

Local tools cover this exercise. Add external connections only for a task that needs them, and inspect their access. MCP server concepts.

Keep study data, exports, and logs outside this workspace. Review study code and images before sharing with an agent; aggregate plots also need a sharing review. See lab data guidance.