Define Done Before Your AI Agent Starts (Or It Never Will)

There's a pattern on Hacker News that keeps coming up: "LLMs work best when the user defines their acceptance criteria first."

It's a great insight for chat prompts. But for autonomous agents, it's 10x more critical.

The Problem

When a human uses an LLM, they course-correct in real time. When an AI agent runs on a cron schedule with no supervision, there's no course-correction. Without defined acceptance criteria, agents either stop too early, run too long, or do the "right" thing in the wrong context.

The Fix: explicit done_when criteria

Every agent in our system has this in its config:

{
  "agent": "content-agent",
  "task": "draft_tweet",
  "done_when": [
    "tweet is under 280 characters",
    "includes a link",
    "hook uses a specific number or metric"
  ],
  "fail_when": [
    "tweet makes promises we can't keep",
    "sentiment is hype rather than educational"
  ],
  "timeout_after": "3 loops"
}

The agent reads this before acting and checks against it before outputting.

Why This Matters

Without acceptance criteria, agents default to plausible completion — they do what looks right. With criteria, they aim at verifiable completion — they do what's measurably right.

The pattern shows up everywhere:

• Consistency: same quality on loop #1 and loop #1,000
• Debuggability: trace failures to specific criteria that weren't met
• Multi-agent handoffs: Agent A knows exactly what Agent B expects

The Three Fields Every Agent Task Needs

• done_when — specific, verifiable conditions
• fail_when — hard stops that override completion
• timeout_after — max loops before escalation

We run 5 agents 24/7 on this pattern. It's what separates "the agent kind of worked" from "the agent reliably works."

Full config templates across 5 agent types → askpatrick.co