ContextRender: keep the tool results your long run actually reuses

Savini Kashmira, Jayanaka L. Dantanarayana, Lingjia Tang, and Jason Mars (University of Michigan) treat long-horizon agent context as a selection problem, not a window-size problem. Passing the full tool-result history to every model call is expensive even when it fits; pruning by recency or semantic relevance can drop identifiers a later step still needs. ContextRender keeps complete tool results in a persistent graph, runs Tool-Flow Analysis to record observed reuse (later model/tool steps that echo values first introduced by an earlier result), and renders a fixed history budget by combining reuse, recency, and embedding relevance—omitted results stay available for later turns.

On AppWorld and 8-objective QA with GPT-4.1, Muse Glimmer 30B, and DeepSeek V4 Pro at a 6K history budget (well below max context), ContextRender beats prune/compact/recency-summary/semantic-retrieval/ACON on task performance in all nine model–benchmark settings, and sits close to or above full history while cutting mean inference cost by 10.2–32.2%. Ablation: adding observed reuse to recency+relevance lifts AppWorld TGC and QA F1 (~4pp relative) and roughly triples recall of results reused more than 25 turns later. Runtime adapters plug into OpenCode, AppWorld ReAct, ACON, OpenClaw, and Meta ARE without rewriting the agent’s tools or grader.

Why it matters for code review

Always-on factories die on context tax as overnight trajectories lengthen: either you pay full-history inference every step, or you drop the contact list / file id / error blob the morning merge still needs. ContextRender’s receipt for continuous harnessed throughput is the same discipline as a merge gate—bind what you show the model to evidence of use, keep the bytes, and re-score before the next call. Score long-running agent loops by later-used result recall under a hard budget, not by how aggressively you compacted the scrollback—see also EA-Graph’s artifact-anchored verification memory and LatentSift’s verification-token cut. FOMO for agents that “keep everything in context” without a reuse graph just burns tokens on history the next step never needed.