Your AI Agents Need an Intent Steward, Not a Better Manager

Your AI agents are building fast. Nobody decided what they're building.

Here's the pattern. You're working with agents on an idea that's still forming. Monday you ask them to index your finished documents. Tuesday, add the dev project folders. Wednesday, the old AI conversations behind that code. Thursday, decades of archived email.

A good execution agent reads that week as four ingestion jobs. Four pipelines, four schemas, clean tickets, visible progress. Nothing about that is wrong. It's just incomplete in the way that matters.

Read the same week from above and a different thing appears: you've moved from indexing finished work to reconstructing how your thinking and decisions developed over time. That's a different product. It wants different structure, different priorities, different trade-offs. The four-pipeline version will work. It will be organised around a purpose you never actually chose.

When the agents get noisy, the instinct is better management. Clearer assignments, tighter monitoring, more check-ins. That's the right answer for work that already exists. It can't help with the moment before the first task, when all you honestly have is a shape. Assign-and-monitor assumes the tasks are real. At that point they aren't yet.

So the missing step isn't a cleaner backlog. It's someone whose job is to hold the unfinished intent without converting it into inventory. Look across the requests, name what they might be becoming, and put that back to the person for a yes or no. Call it an intent steward. It might be a manager, a founder's thinking partner, or a weekly review ritual. It doesn't need to be a new hire.

This isn't a licence to stay vague. The steward's job is to get to a confirmed goal fast. Once you hear yourself say "oh, yes, that's what this is," bring on the schemas, the specs, the tickets and the dig.

The question worth asking before any build starts: is this a settled outcome ready to break down, or a pile of requests whose purpose we haven't understood yet?

Originally posted on LinkedIn