← Back to essays

Put Agents on Full Time and Batch the Human Checks

·3 min read·By Ry Walker

Most companies run their agents part time. The agent does a step, then waits for a person. It drafts, then waits. It fixes, then waits. The human becomes a relay between every action, and the agent's speed disappears into that person's inbox. If you want the speed back, let agents run on their own wherever you can, and batch the human checks that have to stay.[1]

Start by writing down what can go out without a person. In code, we use a simple rule: a change with no schema change, no public API change and no security change can go straight through after automated review. Anything that touches one of those gets a human. I made the case for that gate in the production cut should be a classifier. The rule matters less than the fact that it is written down. Once it is explicit, people can argue about it, sign it and change it, instead of each reviewer guessing.[2]

The same move works in every function. A support team can let agents send answers that match an approved knowledge base and route anything about refunds or legal terms to a person. Finance can let agents reconcile and categorize, and hold anything above a set threshold. Marketing can let agents publish variants of approved copy and stop on new claims. Each team gets a short list of things that always need a human, and everything else runs.

Then batch the checks that remain. When a person has to review, do not make every job stop and wait in line. Let the agents run five or ten jobs at once, then have one person review that set together. It is the difference between stop signs and traffic lights. At a four-way stop, every car halts and goes one at a time, and the intersection crawls. A traffic light lets a whole lane move, then the next one. The intersection and the rules are the same, and far more cars get through. Reviewing a group of results in one sitting works the same way: the reviewer loads the context once, spots patterns across jobs, and sends the whole batch on. It also keeps review from becoming the next wall, the one I described in review throughput, and it fits with moving the standup to the release.

Pick one workflow this month. Write the list of what needs a person, let agents run everything else in the background, and for the work that does need a person, let five to ten jobs finish before someone reviews them together.

Key takeaways

  • Agents that stop for a person after every step are part-time agents, and the person becomes the bottleneck.
  • Write down what can go out without a human, so judgment goes to the work with real blast radius.
  • When a human must review, let agents finish five to ten jobs and review them as a group, like a traffic light instead of a stop sign.

FAQ

What should always get a human check?

Anything with real blast radius. In code that means schema, public API and security changes, and in other functions it means things like pricing, legal terms and anything sent to customers under a new claim.

Does batching reviews make mistakes ship faster?

No. The risky items still get a person, and reviewing them as a group lets the reviewer compare results side by side, which catches patterns one-at-a-time review misses.