Scoped Verification for Reliable Long-Horizon Agentic Context Evolution under Distribution Shift
06:00 · July 13, 2026 · arXiv cs.AI RSS
Deployed LLM agents rely on agentic context, the model-external textual control content assembled by an operational harness. In this work, the mutable component of that context is a persistent system-level instruction that is updated from operational experience while the model, tools, and harness remain fixed. Over long evolution horizons, flat-text maintenance makes verification increasingly difficult as accumulated instructions grow and interact. We propose Graph-Regularized Agentic Context Evolution (GRACE), which maintains the persistent instruction component as a typed semantic graph and validates proposed updates within the local typed neighborhoods of modified nodes. Accepted graph updates are reconstructed as incremental edits to the textual instruction checkpoint used at deployment. We evaluate GRACE within a fixed telecom agent harness derived from $\tau^2$-bench under a controlled distribution-shift protocol. Across five independent replications, GRACE improves strict reliability, measured by pass^3, from the Gemini 2.5 Flash zero-shot value of 0.091 to 0.673$\pm$0.136 at the final checkpoint. This exceeds a Gemini 3.1 Pro zero-shot reference of 0.242 on the same held-out set, while the flat-text HCE baseline finishes at 0.191$\pm$0.051. These results identify two requirements for reliable long-horizon context evolution, a structural substrate that makes verification local and a consolidation mechanism that keeps accumulated instruction content usable.
Summary
GRACE addresses the challenge of maintaining reliable persistent instructions for LLM agents over extended evolution horizons when task distributions shift. In deployed agent systems the mutable control surface is a system-level instruction that accumulates procedural guidance, constraints, and domain assumptions from operational experience while the underlying model, tools, and harness remain unchanged. Flat-text representations make verification progressively harder because relationships among rules become implicit in linear order, increasing the risk of contradictions, context collapse, and degraded reliability as the artifact grows.
The method maintains this instruction component as a typed semantic graph in which atomic units appear as nodes and structural relationships as directed edges. Proposed updates are validated only within the local typed neighborhoods of affected nodes, a form of scoped structural analysis that avoids full-document reprocessing. Accepted graph changes are then materialized as incremental textual edits to the checkpoint supplied at inference time. This substrate therefore separates the representation used for evolution and verification from the flat text consumed by the frozen agent harness.
Evaluation was performed inside a fixed telecom customer-service harness derived from τ²-bench under a controlled alternating distribution-shift protocol spanning ten evolution batches. Across five independent replications, GRACE raised strict reliability (pass³) from the Gemini 2.5 Flash zero-shot baseline of 0.091 to 0.673 ± 0.136 at the final checkpoint, surpassing both a flat-text hierarchical context evolution baseline (0.191 ± 0.051) and a stronger Gemini 3.1 Pro zero-shot reference (0.242). Ablation results isolate the contribution of local structural validation and of a consolidation mechanism that keeps accumulated content usable, confirming that both are required for sustained improvement rather than mere contradiction avoidance.
The work supplies the full codebase, formal definitions of the graph substrate and update operators, and detailed ablation analysis.
Why it matters
High technical depth and novelty in agentic context engineering directly support Dutch AI researchers and advanced practitioners building reliable, verifiable LLM agents for enterprise or regulated domains.
