Adaptive Consistency Graph for Long-Horizon Agents

Adaptive Consistency Graph for Long-Horizon Agents. By Jiecong Wang and two others, from Beihang University and the Qi-Anxin Group. Published on the twenty-sixth of September. Its code repository is worked out: about forty-five source files, with a readme.

Large language model agents often make sensible local decisions on short tasks. Their performance degrades when success depends on long chains of dependent actions and tool calls. Over a long execution, the task requirements, the evidence gathered so far, and the current state can gradually come apart, and once they do, later decisions drift away from the original objective. Nothing in any single step looks wrong. The failure is in the accumulation: each choice is reasonable given what the agent is looking at, but what it is looking at has slowly stopped reflecting what it was asked to do.

The Adaptive Consistency Graph, or ACG, is proposed to keep those pieces connected. It does not replace the base agent's planner or its tool executor. It supplies something to them instead: a structured, traceable view of the context for every decision.

Two layers do the work. A persistent graph grows incrementally, recording execution evidence along with its provenance. Then, for each decision, ACG builds a temporary view centered on the requirements, drawn from that graph and held within a bounded context budget. The graph is long-lived and keeps everything with a record of where it came from. The view is short-lived and keeps only what the current decision needs, organised around the requirements rather than around the order in which things happened.

In the matched evaluation, ACG raises the average success of GPT-5.6-luna from forty-four point five percent with ReAct to fifty point two percent. The largest gain is on BrowseComp-Plus, seventy-three point five percent against sixty-two point four. The authors also analyse trajectory structure and inference cost to characterise where the improvement comes from.

A long task fails quietly when the agent loses the thread between what was asked and what it has found. ACG keeps that thread in a graph and hands each decision a fresh, requirement-first view of it, without touching the agent underneath.