← Back to Feed

Post from 21 October 2026

An AI agent that reasons in text writes out every step as words, including the steps only it will read.

Those text outputs do two jobs. Some are the model’s scratch work, read only by the model. Some are messages to a tool or a person. But theoretically speaking, only the messages have to be words; the scratch work could stay as numbers inside the model.

Generating tokens/words as outputs is computationally expensive, because a model produces text one token at a time and each token takes a full pass through the model. On today’s reasoning models, those tokens are billed as output, even when the API never shows them to you. Think of it like a program that prints every intermediate calculation to the screen and reads it back in.

HybridAgent, from a new research paper, splits the two jobs: the model would reason in vectors (the numerical values it computes internally) and switch to text only to call a tool or give an answer to the user. So it keeps that work in RAM and writes out only for an API call or the final result.

But here’s the tradeoff: reasoning that never becomes words is harder for a person to read.

So if you’re building an agentic workflow using a reasoning model, you might want to check how many of the output tokens you pay for are the model’s own reasoning. You can do that by:

  1. Logging the visible output tokens and the reasoning tokens as two separate numbers for every model call in the loop. Every major API reports the reasoning count in its usage data.
  2. Sum both per task, not per call, as an agent makes many calls per task.
  3. Compare the two. If most of what you pay for per task is reasoning, that’s where the lower-effort setting or a cap on reasoning tokens might cut costs.

Also on Bluesky · Mastodon

Loading comments…