Position: Agentic Evolution is the Path to Evolving LLMs

Abstract: As LLMs move from curated training sets into open-ended real-world environments, a fundamental limitation emerges: static training cannot keep pace with continual deployment environment change. Scaling training-time and inference-time compute improves static capability but does not close this train-deploy gap. We argue that addressing this limitation requires a new scaling axis-evolution. Existing deployment-time adaptation methods, whether parametric fine-tuning or heuristic memory accumulation, lack the strategic agency needed to diagnose failures and produce durable improvements. Our position is that agentic evolution represents the inevitable future of LLM adaptation, elevating evolution itself from a fixed pipeline to an autonomous evolver agent. We instantiate this vision in a general framework, A-Evolve, which treats deployment-time improvement as a deliberate, goal-directed optimization process over persistent system state. We further propose the evolution-scaling hypothesis: the capacity for adaptation scales with the compute allocated to evolution, positioning agentic evolution as a scalable path toward sustained, open-ended adaptation in the real world.

• The paper demonstrates that agentic evolution enables continual deployment-time learning to systematically bridge the LLM train-deploy gap.
• It introduces the A-EVOLVE framework that integrates diagnostic, planning, and verification modules for modular, goal-oriented updates.
• Empirical results show up to +16% improvement in task goal completion, validating evolution-time compute as a scalable resource.

Agentic Evolution as a Scalable Path for LLM Adaptation

Introduction: The Train-Deploy Gap in LLMs

The static performance paradigm—where LLMs are pre-trained or fine-tuned on curated datasets and then deployed—has reached inherent limitations when facing the non-stationary realities of open-ended environments. Static scaling, either at training or during inference, fails to systematically address distribution shifts, recurring failures, confidentiality constraints, and efficiency demands encountered post-deployment. This paper argues that mere scaling of capacity or depth of reasoning does not bridge the "train-deploy gap," which remains the principal bottleneck for robust LLM utilization in dynamic environments (2602.00359).

Formalizing Agentic Evolution: Beyond Static and Heuristic Update Rules

The authors introduce agentic evolution as a third axis of scaling. Evolution here is operationalized as continual deployment-time learning—systematic, cross-episode policy improvement based on accumulated interaction evidence. In contrast to parametric online adaptation (which modifies weights and risks catastrophic forgetting) and non-parametric heuristic adaptation (which appends text or coarsely tweaks prompts), agentic evolution is instantiated as an explicit, autonomous agent (the evolver) that exercises diagnostic, planning, and verification agency over both parametric (LLM weights) and non-parametric (tools, workflows, structured knowledge, test suites) persistent state.

This agentic process obeys three foundational principles:

1. Goal-Oriented Evolution: The evolver actively diagnoses failure, localizes structural gaps, and targets updates to persistent artifacts responsible for observable errors, transforming repeated mistakes into governed, reusable competence.
2. Autonomy: Updates are conditional and marshaled through decision-theoretic validation gates, allocating evolution compute only when changes are necessary and beneficial, thereby maintaining system stability.
3. Compositionality: Evolution refines modular artifact registries—tools, skills, validation assets—that are explicitly verified and versioned, enabling recovery, auditability, and preventing context saturation typical of passive memory accumulation.

These principles are formalized in the A-EVOLVE framework, which operationalizes the solve/evolve loop, persistent artifact registries, structured edit operators, and separated compute budgets for online evolution versus inference.

Empirical Results: Effectiveness, Reliability, and Scaling of Agentic Evolution

Experimental evaluation is performed on the AppWorld environment, a benchmark for interactive coding agents with strict verification and complex tool utilization requirements. Two principal metrics are reported:

• Task Goal Completion (TGC): Fraction of tasks fully solved.
• Average Passed Tests (APT): Fractional completion per task, reflecting incremental improvement.

Key Findings

• Capability Multiplication: Agentic evolution delivers substantial improvements across all solver backbones (up to +16% TGC over state-of-the-art non-agentic baselines, and matched or exceeded larger vanilla LLMs with smaller but evolved models), indicating reliable conversion of deployment evidence into durable capability.
• Component Contributions: Ablation studies confirm that each loop module (diagnosis, analysis, planning, verification) contributes synergistically; removal of verification, in particular, yields the greatest instability.
• Evolution-Scaling Hypothesis: Increasing evolution-time compute (either via iteration steps or evolver backbone size) yields monotonically increasing attainment of the performance frontier, in contrast to early saturation seen in heuristic approaches. Larger evolvers produce higher-quality diagnoses, more robust artifacts, and fewer verification failures, converting compute resources into sustainable capability more efficiently.

Theoretical Implications: A New Scaling Regime for Adaptive Intelligence

The evolution-scaling hypothesis is analytically characterized: the performance ceiling of a deployed LLM system ($P^*(C_{\mathrm{evolve}}, T_0)$) is strictly increasing in evolution-time compute, holding solve-time compute fixed. The implication is that adaptive intelligence is not a stochastic by-product of opportunistic fixes, but can be engineered as a scalable optimization process governed by the evolver’s competence and allocated compute. This view formalizes adaptation as convergent, persistent improvement, and distinguishes resource allocation for immediate problem-solving versus systemic evolution.

Practical and Societal Impact

Agentic evolution enables local, privacy-preserving adaptation through artifact evolution at the deployment edge, alleviating the need for centralized retraining that risks privacy leakage. Persistent artifact synthesis amortizes one-shot inference costs into reusable and verifiable capabilities, promoting compute sustainability in long-horizon deployments. To mitigate risks of drift or unaligned optimization, the framework incorporates governance gates and modular verification, ensuring interpretable, auditable changes.

Critique of Alternative Paradigms

• Inference-Time Depth Alone Is Insufficient: Extended reasoning only amortizes for novel instances; agentic evolution compiles repeated patterns into persistent systems assets.
• Blind Parametric Plasticity Is Unstable: Direct online weight adaptation challenges interpretability and safety due to catastrophic forgetting and opaque update provenance.
• Heuristic Evolution Plateaus Early: While lightweight, heuristic memory or prompt accumulation lacks strategic updating and reasoning, leading to saturation and diminishing returns.

Future Directions

Potential lines of progress include:

• Benchmarking agentic evolution in even more entropic and non-stationary environments, measuring durability and reuse of evolved artifacts.
• Framework refinement for more efficient diagnosis, compositional planning, and robust verification in heterogeneous environments.
• Formal theoretical characterization of regret bounds, separation results, and resource-optimal adaptation laws in evolving program spaces.

Conclusion

The argument advanced positions agentic evolution, instantiated as autonomous evolver agents with structured artifact updating, as a scalable, pragmatic path for adaptive LLM deployment in real-world settings. By introducing a new axis of scaling—evolution-time compute—the work provides an operational and theoretical foundation for moving beyond static inference, toward systems that reliably, autonomously, and safely close the train-deploy gap with systematic, persistent capability improvement.

Reference: "Position: Agentic Evolution is the Path to Evolving LLMs" (2602.00359)