Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 5 min read

Most of our Claude spend went on agents talking to each other

In one week, 57% of our Claude spend went on turns started by another agent's message. Here is how we got there, what else broke along the way, and what we changed. In short From 1 to 8 November, 57% of our Claude spe

In one week, 57% of our Claude spend went on turns started by another agent's message. Here is how we got there, what else broke along the way, and what we changed.

In short

  • From 1 to 8 November, 57% of our Claude spend went on turns started by a message from another agent. About 20% went on turns a person started.
  • Two things had already gone wrong on the way: a workflow engine we ended up deleting, and a slowdown we first put down to hardware.
  • One fix has shipped: agents take shared machine capacity from a lock instead of asking another agent for it. It is too early to measure.

We didn't know whether we were efficient

We run a lot of AI coding agents, and for a long time we didn't know whether we were efficient with our tokens. We just kept coding. At the start the idea was to churn out products and see how far AI coding could take us. As the workload grew and we tried to build more, that stopped being enough.

So we asked the agents themselves to look into where the spending went.

For the week of 1 to 8 November, the answer was this. 57% of our Claude spend went on turns started by another agent's message, not by a person. The median one of those turns landed on a context of about 500,000 tokens. People started about 20% of the spend. (That is Claude spend only, counted in weighted token units, not dollars.)

I wouldn't say the number surprised me, and I had no guess to compare it with. While you are using AI agents, you cannot tell whether they are efficient. You just get the work done. The only thing you notice naturally is speed. Sometimes, from the way an agent responds, you get a feeling that things are going in the wrong loop. But that only happens if you are paying attention, and a feeling is not objective and not a number.

Why there are so many agents in the first place

I started with plain command-line terminals, running Claude Code and Codex. I wanted to learn fast and build a lot, so I was working with more than ten sessions at once. Plain terminals give you no proper workspaces, no alerts, no way to resume and nothing to sort sessions by. I remember being afraid to restart my computer, because I would lose every session and have to bring them back one by one.

That is why SoluTerminal exists. I built it for myself, my team started using it later, and now anything I build with AI is built in it.

A normal day looks like this. Each project has its own workspace. Inside it, different agents hold different roles: one for security, one for pushing releases, and a lead. I mostly talk to the leads, and the leads can start their own teams of worker agents. My job is the parts that still need a person: an intervention, an approval, something that has to be connected by hand. Every release still waits for a person's own typed approval.

The agents talk to each other by design

Agents on different models, Claude and Codex, can message each other across sessions. When we first opened that up, we set very few guardrails and gave very few instructions. We told them, in effect: you can talk to this agent, you can talk to that one, go and coordinate and get it done.

They did. In eight days they sent each other about 7,800 messages. Another 1,457 went through a tool our own rules said not to use as a messenger. Some agents were simply repeating things to other agents, which is totally unnecessary.

What broke first: a workflow engine

The open approach was itself a replacement. Before it, we tried something more structured.

Different models have different strengths and weaknesses. The idea was a workflow that used each model where it was strong and kept the cost down, instead of using the most intelligent model for every task. It was ambitious: many roles, many stages, many hand-offs. With that many moving parts, one failure causes others. If one model ran out of quota, the workflow broke.

I wouldn't say it failed completely, or that it can never work. But every task took significantly longer, and I had to step in by hand to look after the parts that had broken.

The record agrees. In October the engine landed 0 of 69 tasks, while 195 pull requests merged through direct agent sessions. We deleted it: about 532,000 lines, about half of them tests. The decision record said "delete later". It was deleted the next day.

I knew it wasn't working when I saw how much time the engine itself was taking, and how far it was delaying everything else. Deleting it was fine. It hurts only because of the time I spent on it. If I had moved away from it earlier, I would have had that time and those tokens back, and more would have been built.

What broke next: a slowdown that felt like hardware

With the engine gone, things went back to normal for a while. Then running tests became the problem. My computer could not handle that many heavy jobs at once, and work had to queue. We moved heavy jobs to cloud machines, and things moved again.

It felt like hardware. The numbers showed a second cause. The automatic checks took 11 minutes the whole time. Yet merges fell from 33 a day to about 12 (4 to 10 November, against 24 October to 3 November), and the middle time from first commit to merge went from 1.1 hours to 13.6 hours. An internal review blamed the ceremony around the code: agents waiting for another agent to grant a "go", and a main branch that was frozen for about 37% of hours. The same week also had a disk incident and a filming day, so the picture is not perfectly clean.

We acted on both causes.

What we changed

  • Heavy jobs run on cloud machines, not on one computer.
  • Agents take shared machine capacity from a lock, not from a message to another agent. This shipped on 11 November. It is one fix, and it is too early to measure.
  • We keep reviewing how agents talk to each other, so tokens are not spent on agents repeating themselves, while we keep the benefits of coordination.

What I take from it

Efficiency has to be part of the game. In practice that means prioritising: making time to review how the work is actually being done, instead of brute-forcing through everything and spending more tokens to reach the goal. And it means reviewing often enough to decide when to move away from something.

We build SoluTerminal, an open-source Windows app for running Claude Code and Codex side by side. It's free during beta. https://soluterminal.com/?utm_source=devto&utm_campaign=oss-launch&utm_content=agents-article

If you run more than two agents at once, what's your rule for when they're allowed to talk to each other?

Drafted with AI assistance from our team's notes; checked and edited by a person before publishing.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.