YLEM
Fields / 04

AI Coding

Coding is the first domain where agents do substantial delegated work. The bottleneck has moved from writing code to reviewing and verifying it.

DeployedAcceleratingRevised 5 Oct 2026Draft — awaiting editor review
What to watch
  1. How review, testing and CI adapt to much higher code volume
  2. Agents working in the background, in parallel, on many tasks
  3. Effects on how engineers learn and how teams are structured

The state of play

Coding moved through three phases quickly: inline completion (Copilot, 2021), chat inside the editor, and agents that read a repository, run commands and open pull requests. Code is an unusually good domain for AI because the environment answers back: compilers, type checkers and tests give fast, objective feedback.

What works today

  • Implementing well-specified features and fixes in existing codebases.
  • Writing tests, migrations, refactors and glue code.
  • Explaining unfamiliar code and investigating bugs.

What doesn’t yet

  • Ambiguous product decisions. Agents implement the spec, including a wrong one.
  • Large architectural changes without a human holding the design.
  • Avoiding subtle, plausible-looking errors that pass weak tests.

Key ideas

  • Verification is the product. Teams with good tests get far more out of agents.
  • Context engineering. Repository instructions, conventions and examples shape output more than prompts do.
  • Review load. More generated code means review, not writing, sets the pace.

Open questions

What does a junior engineer’s path look like when agents do most junior-level tasks?