What Developers Should Start Giving AI to Code

The unglamorous, high-volume work

Migrations, deprecation cleanups, and mechanical refactors across a large codebase are exactly the kind of work most engineers put off because it is tedious rather than hard, and exactly the kind of work an agent can grind through consistently, checking each change against tests as it goes.

These tasks are underused not because agents perform poorly on them but because they feel too mundane to delegate deliberately. Engineers reach for AI on flashy new-feature work and forget it is even better suited to the backlog of cleanup nobody wanted to do by hand.

Test and documentation debt

Backfilling test coverage on old, stable code and writing or updating documentation that has drifted from the actual implementation are both strong fits, since the agent can read the current behavior directly and describe or verify it, rather than needing to invent new behavior from an ambiguous spec.

This kind of work compounds. Every piece of debt paid down here makes the next task, AI-assisted or not, easier, because both future agents and future engineers benefit from better tests and accurate docs.

• Large mechanical refactors and API migrations
• Backfilling tests on stable, well-understood code
• Updating documentation to match current behavior
• First drafts of boilerplate for a well-established pattern
• Investigating and summarizing an unfamiliar part of the codebase

Investigation, not just implementation

Beyond writing code, agents are underused for investigation: tracing how a specific behavior actually works across a large system, or summarizing what changed in a dependency's changelog that might explain a new bug. This kind of research work benefits from an agent's ability to read quickly across many files without getting tired.

The common thread across all of these is that they are tasks with a clear, checkable outcome and low ambiguity, exactly where agents perform best, and exactly the category of work most teams have a large, unaddressed backlog of.