Goal
The verifiable objective given to the agent for a task. It is different from the outcome, and the system breaks when the two become misaligned.
In one sentence
The operational objective given to the agent: a specific, verifiable instruction for what it must accomplish in a task. It is distinct from the outcome, and agentic systems often break when the two become misaligned.
Before
Outcome — not verifiable
What it does
Goal — specific and verifiable
After
The agent knows when it is done
Example
Outcome: sponsor who renews. Agent goal: write a 180- to 320-word letter, in the child’s voice, where every statement comes from the field report.
The goal is verifiable and the outcome is not, so the two need to exist separately. The project risk lies in the space between them.
The common mistake
Giving the outcome to the agent as if it were the goal: "make the sponsor happy". It has no way to verify that, so it invents its own criterion, usually the easiest one to satisfy. You get effusive, empty letters, and they pass the test the agent chose for itself.
In practice
- Write the goal together with the verification criterion. One without the other is not enough.
- Pass the stranger test: can an outsider tell whether it was achieved, without asking anything?
- Version it together with the prompt, because the two change and break together.
How to make it tangible
The instruction that goes into the agent’s prompt, with verifiable criteria: format, source, limits. It lives inside the spec, versioned.
Connects with
Translated from Portuguese with AI assistance.