Brief
OdoBot cuts web-agent token use by 44-80% on a 45-task test
A new web-agent architecture called OdoBot completes tasks with fewer tokens than two state-of-the-art competitors, according to a paper on arXiv. It uses a behavioral model built from successful task demonstrations.
The paper introduces OdoBot, a web-agent architecture that analyzes an application's user interface and interacts with it to automate tasks described in natural language. Instead of relying solely on live UI analysis, OdoBot builds a behavioral model of the application from successful task-execution demonstrations. In experiments with 45 tasks on the Canvas Learning Management System, OdoBot used 44% fewer tokens than Agent-E and 80% fewer than WebVoyager. It also surpassed WebVoyager in task success rate. The source does not provide absolute token counts or success rates.
Our reading
Our reading is that this approach could make web automation more affordable for routine tasks, though the evidence is limited to one application and a single set of experiments.
What to do or watch
Watch for replication of OdoBot's token savings and success-rate results beyond the single Canvas LMS experiment, since the paper reports only 45 tasks on one application. The precise unresolved question is how OdoBot compares with Agent-E on task success rate, and what the absolute token counts and success rates are, which the source does not provide.
Source details and supporting facts
Each line is stated by the page named above it.
Stated by arXiv
- OdoBot uses 44% fewer tokens than Agent-E.
- OdoBot uses 80% fewer tokens than WebVoyager.
- OdoBot surpasses WebVoyager in task success rate.
- Experiments involved 45 tasks on the Canvas Learning Management System.
- OdoBot builds a behavioral model from successful task-execution demonstrations.
Sources
- arXivText stored 16 September 2026
How this story was checked. Written from the 1 page listed above, stored 16 September 2026; claims checked against that stored text on 16 September 2026.
What that means
- 5 of 5 reported statements were confirmed against the page that carries them; the rest were removed rather than published.
- Figures in the text were required to appear in the stored source text: yes. Identifiers: yes.
- The check reads stored text only: no claim rests on a fresh look that did not happen.
- Where the reporting was silent, the text says so instead of filling the gap.