Yutao Sun and colleagues submitted a preprint on 7 November 2026 reporting that executable finite-state control improves language-model agents’ adherence to procedures over long runs. Their proposed finite-state machine (FSM) stores workflow state outside the model’s context and enforces the procedure at tool calls, state transitions and the end of an agent’s turn.[1]
The authors say agents drift from procedures supplied at the start, and that repeating instructions or resuming a stopped agent merely shifts the shortcut they take. In an experiment described as pre-registered, using 200-account SOP-Bench batches, Claude Haiku 4.5 with the FSM averaged 139 accounts decided correctly and according to procedure, compared with 48 for the strongest baseline. The abstract reports a 95% confidence interval of [46, 130] for the difference and says performance persisted through the end of the run.[1]
The authors also report that Haiku 4.5 with the FSM averaged more procedurally correct decisions than plain Claude Sonnet 4.6 at 38% of its cost. They describe additional analyses on SWE-bench Verified, τ²-bench, and staged and cross-context workflows as suggesting benefits from stronger execution control. These are research claims reported in the abstract.[1]
[1]: https://arxiv.org/abs/2611.10884
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
71a036116c8909f9438108a6b708c818aead9e7e922c8846bdee44df2a85f357