CoursesAgent Harnesses in Practice
Tools and Safe Edits
How a harness makes an edit safe to trust: anchors that prove the file is unchanged, a language server and a debugger that hand back evidence, a permission table where order decides, and subagents that return an object you can check.
- Lessons
- 6
- Exercises
- 34
- Minutes
- 36
- 1
Edits That Prove Freshness
After this lesson you can explain why an edit must prove the file is unchanged, read a rejected edit as good news, and spot the search-and-replace loop that ends in a silent overwrite.
DebugMultiple choiceTrace - 2
The Editor in the Loop
After this lesson you can say what evidence comes back to an agent after a write, why a diagnostic beats "it looks right", and which renames and debugging steps a language server and a debugger make safe.
DebugFill blankMultiple choice - 3
Permission Tables
After this lesson you can read an OpenCode permission block, say which rule wins for a given command, fix a block where the order lets a push through, and say what --auto changes.
DebugFill blankMultiple choice - 4
Typed Subagents
After this lesson you can say what comes back from a subagent and what stays in its own window, pick the built-in subagent for a job, and explain why a typed return is checkable and a prose return is not.
DebugMultiple choiceTrace - 5
Checkpoint: Tools and Safe EditsCheckpoint
Stale anchors, diagnostics, permission order and typed returns, in fresh situations.
DebugCode orderFill blankMultiple choice - 6
Boss: The Risky RefactorBoss
Ana renames src/parcel/loader.ts to parcel-loader.ts across shipwise while Tomás keeps editing the same file. Read the permission block, catch the rule that lets a push through and fix its order, handle the stale-anchor rejection, act on the diagnostic after the rename, check the subagents' typed reports, and hand off with evidence.
DebugCode orderMultiple choiceTrace