agentbelt 0.3.0: token optimization for local AI agents, with honest numbers

The latest version of agentbelt, 0.3.0, introduces token optimization for local AI agents using the Goose adapter and a demo GIF. To test the improvements, the author ran the same tasks with compaction turned off and on, conducting three runs each. The token counts were measured using Ollama reports.

The largest prompt read five files both with and without compaction. However, the largest prompt was shorter in every run with compaction enabled. On a log-heavy task, input tokens were reduced by 56%, and the task finished successfully in all runs. In contrast, the file-reading task showed no average token savings, and one run even had to re-read files, costing more.

The author notes that the biggest benefit for local agents is avoiding repetition of the same logs. Overly shortening file contents can lead to issues where the model re-reads lost information. To try agentbelt, run `npx agentbelt setup` and then instruct your AI app to use `agentbelt`'s `agent_run` with `compact: true`. For those who find the tool useful, the author encourages a star on GitHub and requests feedback on any measurements taken while running agents on local models.