Everyone assumes the next phase of AI engineering is cost optimization. Squeeze the prompts, cache the context, cut the token bill. I think that instinct is exactly backwards. Most teams have not yet reached the token spend that real engineering work requires. They are underspending, and the underspend shows up as mediocre pull requests.
The one-shot generation is genuinely impressive now. But an agent can go down a wrong path in the middle of a session and never realize it, because nothing in the session pushes back. The fix is not a better prompt. The fix is a loop, an actor-critic structure where two or three different coding agents draft, attack, and revise the same unit of work before a human ever sees it. We run these loops across multiple agent products at once, and yes, it burns more tokens than trying to one-shot the task.
But here is the accounting most teams skip. Quality on a PR has a fixed price, and you pay it somewhere. Either you pay in tokens, letting agents iterate against each other in the background, or you pay in senior engineer hours during review, which is the most expensive compute you own. When you let the agents fight each other first, the human gets a PR that has already survived an adversarial pass.
The loop is not free lunch, to be clear. A critic can drag the work further off course, padding the diff with irrelevant tests or inventing requirements. That is why the loop needs a harness, defined roles, and a stopping condition. This is software engineering, not a group chat.
The token reckoning is coming, but it will not reward the teams that spent the least. It will reward the teams that figured out where spend converts to quality and pushed budget there deliberately. Cheap and wrong is still wrong. Spend more, in the right loop, and the review burden downstream collapses.
Key takeaways
- A single agent in a single session can go down a wrong path and never know it, which is why one-shot generation caps out on quality.
- Running multiple coding agents in an actor-critic loop costs more tokens but produces materially better pull requests.
- You pay the quality price on every PR one way or another, either in tokens up front or in human correction later.
FAQ
Isn't running two or three coding agents on the same task wasteful?
It costs more than a single one-shot attempt, but quality has a fixed price. Either you spend tokens letting agents critique and iterate on each other's work, or you spend expensive human hours catching what a single pass missed. The multi-agent loop is usually the cheaper of the two.
Do critic agents always improve the output?
No. A critic can pull the work further off course, for example by demanding irrelevant tests. That is why the loop needs structure and a defined stopping point, not just more agents talking to each other.
Related Essays
The Token Reckoning Is Coming
Engineers burn ten to fifteen million tokens a day. Powerful models run tasks that do not need them. By the end of 2026, the CFO will start asking questions and most teams will not have answers.
Own the Compute Layer or Pay Someone Else's Margin Forever
Building your own sandbox infrastructure is painful, but it gives you pricing power, feature velocity, and margins your competitors will never match.
Running Out of Tokens Is a Buy Signal
When engineering teams hit their token caps, that is not a cost problem to contain. It is the clearest signal you have that AI adoption is actually working.
Drowning in pull requests that need your review? Try Tembo Review, a beautiful AI-assisted PR review tool unlike anything you’ve used.