Give a long-horizon coding agent an explicit execution ledger instead of a longer history: Pass@1 56.2%→64.2% at 28.9% lower cost, with zero extra LLM calls

arXiv 2608.00808·high signalXBlueskyLinkedInCopy link

Ledger is a deterministic runtime wrapper that distills an agent's completed interactions into explicit state — what has been observed, what has been modified, what has been attempted — then applies it at two boundaries: an 'inform' path that appends a compact state view to the prompt, and a 'govern' path that checks each proposed command against the ledger, returning still-valid earlier results instead of re-executing and flagging redundant repetition. Across all 500 SWE-bench Verified instances it lifts Pass@1 from 56.2% to 64.2% (GPT-5 mini) and 75.8% to 81.0% (MiniMax M2.5) while cutting total cost 28.9% and 31.8%; attached to OpenAI Codex it adds 3.4 points at 24.4% lower cost. Ablations credit most of the accuracy gain to govern and most of the savings to inform — the practical lesson is that agents don't need a shorter view of history, they need a machine-maintained account of their own execution state, which you can bolt on without touching the agent.