Selective revalidation improves decision accuracy for long AI tasks
From Version Conflicts to Decision Conflicts: Selective Revalidation for Long-Running AI Agents
Artificial IntelligenceDatabases
Summary
Long-running AI agents sometimes plan actions based on information that changes before the action happens, causing problems if the action is no longer justified. The authors describe a method called ATR that keeps track of the exact conditions needed for an action and rechecks only what changed before allowing the action. This helps avoid unnecessary blocks or mistakes by focusing on what really matters, improving decision reliability without checking everything again. Their tests showed ATR works well in many scenarios, running faster and more precisely than full rescans, though it is not yet ready for all real-world uses.
AI agentoptimistic concurrency controlversion conflictdecision conflictrevalidationexternal actioncondition checkingtransactionSQLiteperformance evaluation
Authors
Yongjian Lyu, Yang Ren, Ruofei Lai, Wenting Liu
Abstract
Long-running AI agents may read state, reason, wait for tools or human approval, and perform an external action much later. The state that justified the action can change in the meantime. For example, after an agent proposes an 80 GBP refund under a limit of 100, a customer-name change affects only presentation metadata, a new limit of 90 still permits the refund, a limit of 50 invalidates it, and a refund issued by another worker must prevent a duplicate. Standard optimistic concurrency control and version checks can detect that previously read state has changed, but by themselves do not determine whether that change invalidates the pending action's justification. We call any detected version change a version conflict; when that change invalidates the action's justification, it is also a decision conflict. ATR records the explicit, executable conditions that justify a pending action and rechecks only the conditions affected by a change before releasing the external operation. It can retain the action, refresh non-decisive metadata, require replanning, or block execution; a target-side transaction or compare-and-set binds checked state to commit. Across 210,000 controlled executions over 15 mutation cases, ATR matched every developer-specified outcome with no false allows or blocks. In ten durable SQLite checkpoint/resume cells, it evaluated 0.6 conditions per change versus 6.0 for FullScan. At 4,093 recorded reads, ATR took 9.3 microseconds versus 2595.9 microseconds for FullScan. These deterministic results establish controlled feasibility, not production generality or automatic extraction of the required conditions.