Working with Coding Agents.

The delegate-and-supervise discipline of working with autonomous coding agents: when to delegate, scoping work, spec-driven development, instruction files, managing long-run context, reviewing agent-authored PRs, agent failure modes, permissions/sandboxing, verification & accountability, fleet supervision, and honest impact measurement. Distinct from ai_dev_workflow's in-loop AI-assisted craft (boundary = unit of delegation). Covers 11 sub-skills.

• Working with Coding Agents · 1 of 3

  Spec-Driven Development

  In an agent workflow, where is the least expensive point to inject human judgment and steer the work onto a new track if needed?

  AAfter the change has merged, during the next code review on the pull requestBAt the very start, by writing an extremely detailed prompt up frontCOnly once the entire diff is fully ready, by reviewing every changed line closelyDRight after the model proposes its plan, before it writes the implementation

• Working with Coding Agents · 2 of 3

  Agent Failure Modes

  Asked to fix one null check in auth.ts, an agent returns a 40-file diff that also renames variables and reformats several unrelated modules. Which behavioral failure is this?

  AConfident fabrication of an import for a module that does not exist yetBReward hacking aimed squarely at the repository's lint formatting scoreCSilent scope creep well beyond the narrow change that was requestedDPremature memory compaction that discarded the original instruction

• Working with Coding Agents · 3 of 3

  Agent Failure Modes

  While fixing a build error an agent inserts a call to array.flattenDeep() in plain JavaScript and assures you the method is standard. It does not exist. Which failure is this?

  AA version skew between the lockfile and the actually installed packageBA tree-shaking pass that pruned the referenced helper from the bundleCA transpilation target silently dropping a genuinely supported builtinDA confidently asserted hallucination of a function that is not real