Can you make an AI agent better at remembering what matters, just by changing the system around it?

This explores whether you can improve how an agent handles its context window by changing the system around the model instead of its weights. The corpus says yes, and there are several distinct ways to do it. The most direct is to give a frozen agent an outside context manager. Can an external manager handle context for frozen agents? trains a separate RL-based manager to decide what to prune and what to preserve, while the agent itself never changes. The manager has to be tuned to the agent it serves. Stronger agents do better when context is preserved in high fidelity, and weaker agents need aggressive compression. So good context management is specific to the agent, not a universal recipe.

You can also improve it in the harness, the scaffolding around the model, with no learned component at all. Can agent harnesses be automatically optimized across many environments? ran automated search across many environments and found four mechanisms (action execution, context compaction, observation handling, delegated reading). Together they cut token traffic by roughly 45–49% with comparable performance on a 51-task benchmark. [[harness-scaling-lifts-fixed-models-on-terminal-bench-21-without-touching-weights]} shows the same idea on accuracy. Optimizing the execution system around frozen weights lifted several models on Terminal-Bench 2.1, and the same runbook carried over to newer models unchanged. That suggests harness gains and model gains add together instead of competing.

A third route is to stop treating the context window as the place where the task's state lives. Can task state management alone improve long-horizon agent performance? moved task state out of the executor and checked it with independent environment audits instead of trusting the agent's own claims. Qwen 3.7-Plus went from 51.8% to 80.7% on WeaveBench with no model change. Memory works similarly. Can agents learn continuously from experience without updating weights? has agents improve through case, subtask and tool memories rather than gradient updates, reaching 87.88% on GAIA validation. Can agents reconstruct memory on demand instead of retrieving it? goes further and rebuilds memory on demand by traversing a graph as evidence accumulates. It gains up to 23% on reasoning tasks and costs fewer tokens than a fixed retrieve-then-reason pipeline.

The cheapest lever is often the wording and structure of the context itself. Can context quality alone predict how agents will behave? held the model fixed and varied only the context. Scores across seven criteria (role clarity, guardrails, instruction consistency, tool schemas, grounding, injection hardening, token efficiency) predicted matching behavioral successes and failures. So you can audit context quality before you ever run the agent. If you also want to adapt while deployed, Can agents adapt without pausing service to users? describes injecting skills from failures within seconds and with zero downtime, and saves the slower weight updates for idle windows.

The limit is that the most active form of context management, learning to delegate subtasks to subagents and integrate their summaries, is trained. Can delegation teach models to manage context more actively? found that delegation training beats passive compression and transfers to single-agent tasks, which suggests it teaches disciplined decomposition, not just orchestration. So the no-retraining options above mostly get you better bookkeeping (what to keep, where state lives, what to re-read). Learned delegation may be what gets you an agent that manages its own attention.