Cutting agent token costs by reasoning without long text traces
A 's written may not faithfully show its real internal computation. It can produce convincing steps and still reach a wrong answer, or show messy steps while reaching the right one, so the trace is not a dependable audit record. Generating intermediate work one token at a time also increases delay, cost, and context use.
Coconut explores doing intermediate reasoning in latent space and turning it into language only at the end. HRM and HRM Text separate slower planning from faster , while RecursiveMAS lets agents exchange instead of long text messages. The central choice is whether language should be an interface for communication or the medium used for every part of the computation.
Moving reasoning out of readable text may save tokens, but it also makes the model's decisions harder to inspect and deepens the black-box problem.
Key points
- A written may not match the model's actual internal computation.
- Producing long increases token cost, delay, and context use.
- Coconut performs intermediate reasoning in latent space and produces language at the end.
- RecursiveMAS replaces long messages between agents with .
- Lower token use may come at the cost of a larger black-box problem.