The New Bottleneck Isn't Typing, It's Deciding What to Delegate

Why it matters: When an agent can write the code in seconds, the engineer's value moves entirely to the choices made before it starts, and teams that keep measuring output by lines written will fall behind fast.

The Details:

• Designing an agent means setting its permissions, not just its prompt. A prompt shapes one exchange; an agent needs defined tools, explicit boundaries on what it can touch, and rules for how it recovers from failure. Withholding a capability, like the power to edit production code, does more enforcement work than any instruction telling it to behave.

• Missing context turns agents into confident repeat offenders. An agent investigating a flaky test needs to know what counts as flaky in that specific codebase and what's already been ruled out, or it will re-diagnose the same stale bug indefinitely while appearing productive the whole time.

• Vague framing produces technically correct, practically wrong results. Told to "fix the failing test," an agent may simply rewrite the assertion until it passes. The fix against this is writing down the invariant, what must remain true, before handing off the task, not just describing the desired outcome.

• Supervision has to be scheduled, not continuous. Watching every action recreates the bottleneck agents were supposed to remove. Instead, engineers now pick fixed checkpoints, after a plan, after a diff, before a commit, and stay alert to trust drift, the point where early good runs make later reviews sloppier right when something important breaks.

Bottom Line: The engineer's real job now happens before the agent starts working, not while it runs.