← BlogBen LaiSubscribe via RSS ↗

The /goal command is a confession

Coding agents now ship with explicit "set a goal" commands. Last year's pitch was that they figured out the goal themselves. The walking back is real, and worth understanding.

Look at the latest releases of the major coding agents. They all added some variant of /goal or /objective or /plan-first in the last six months. The marketing language for these commands is upbeat — “structured work,” “explicit alignment,” “goal-oriented agents.” The actual reason they exist is less flattering: the agents don’t know when they’re done.

A year ago the pitch was that you would describe a problem and the agent would decompose, plan, execute, and stop. The first part worked. The last part — knowing when to stop — turned out to be the unsolved one. So you got loops. Refactors of refactors. Helper functions for helper functions. Three days of credit burn on what was a thirty-minute task.

Why the goal command is good design

Because it makes the failure mode explicit. The agent will only stop when it can answer “did I do the thing the user asked for?” — and that question requires the user to write the thing down in a stable, referenceable place that the agent can return to. That’s what /goal is. A spec the agent can compare its work against.

If you treat the goal command as a workaround, you’ll skip it and pay for it in tokens. If you treat it as an interface, you’ll start writing better goals — and your reviews of AI work get cheaper too, because you have a contract to compare the diff against.

The bigger pattern

A lot of “agent intelligence” wins of the last two years are actually wins on stop-condition design. Tool-call limits. Iteration caps. Cost ceilings. Required user confirmations. These look like guardrails. They are more like grammar — they let the agent say a sentence instead of a run-on paragraph.

The agent that asks you for a goal isn’t admitting weakness. It’s admitting that the goal is the part that has to come from you, and the rest will follow.