DEV Community

qing zhang
qing zhang

Posted on

How I Automated My Workflow Using ChatGPT Agents — By Fixing One Critical Failure Mode

#ai

Most AI-automation posts talk about “agents doing everything.”
Mine didn’t work—until I fixed one microscopic but fatal detail:

Forcing the agent to say “I don’t know.”

This single constraint changed the entire system.

The Problem: Agents Guessing

When my workflow automation first went live, the planning agent kept inventing missing context:

wrong deadlines

imaginary stakeholders

nonexistent dependencies

Not malicious—just probabilistic.

In automation, a confident guess is more dangerous than a clear error.

The Fix: A Strict “Blocker-First” Contract

I rewrote the agent’s system prompt around one rule:

If any required field is missing, return:
{
"blockers": [
{"field": "X", "question": "What is the deadline?"}
]
}
Do NOT infer. Do NOT approximate. Do NOT proceed.

And I enforced JSON validation at the orchestration layer.

If the agent guessed → schema failed → task dropped.

If the agent asked a question → I answered once → pipeline resumed.

Why This Works

This turns the model from a “predictor” into a deterministic validator.

No silent assumptions

No phantom tasks

No hallucinated requirements

No speculative routing

You force the agent into a request-for-information loop,
which mirrors how real engineers block work when specs are incomplete.

Result

Error rate dropped from 27% → 0% in 48 hours.
Suddenly, my automation became reliable:

specs were correct

release notes matched reality

planning tasks stopped drifting

nothing moved forward without defined inputs

A single constraint unlocked the entire workflow.

Top comments (0)