An AI agent that runs for minutes or hours collects a lot of material: documents, tool outputs, intermediate results, failed attempts. Everything it keeps in its context window costs compute, and everything it drops may be missing later. How an agent manages this context increasingly decides both answer quality and operating cost. New research on Context Language Models (CLMs) suggests that models can learn to manage their own context, with notable gains in accuracy and efficiency.
With every step, an agent's context grows. Long contexts raise compute costs, slow responses and bury the important details among irrelevant ones. Most systems today handle this with one of three fixed approaches, each with known weaknesses:
All three depend on rules that humans designed in advance. That works for predictable tasks, but long-running agents often meet situations those rules did not anticipate.
In a recent paper, researchers from Meta, MIT and the University of Washington describe models that treat their own context like a file they can edit. A CLM can rewrite older messages, keep important facts, delete irrelevant intermediate results, maintain progress notes and track which approaches have already failed. Users can also steer this behaviour by describing the strategy they want.
The researchers tested three set-ups: zero-shot, without special training; in-context learning, where the model refines its strategy from natural-language instructions; and reinforcement learning, where the model is trained on task success with compute efficiency as an added goal.
The reported results, as stated in the paper:
|
Set-up and benchmark |
Reported result |
Note |
|
Zero-shot, BrowseComp-Plus |
11.4% higher accuracy, 21.5% fewer FLOPs |
Compared with the strongest baseline |
|
Zero-shot, 12-hour EdgeBench |
5% higher scores, 59% fewer FLOPs |
Long-running task |
|
In-context learning, ContextBench |
Up to 35.9 points higher accuracy |
Synthetic diagnostic suite |
|
Reinforcement learning, Qwen3.5-9B |
Accuracy from 28.8% to 42.5%, 12% fewer FLOPs |
BrowseComp-Plus |
Two caveats matter for business readers. The FLOPs figures are theoretical compute estimates, not measured runtime, and the results come from specific benchmark tasks. They show a promising direction, not proven production performance.
The researchers and early commentators also point to limits that matter for enterprise use:
CLMs are still research. The underlying lesson already applies: context is a design decision, not a by-product. Five practical steps:
Hypothetical scenario: a mid-sized logistics company uses an agent to process customer complaints. Each case includes long e-mail threads, shipment data and contract terms. Early on, the agent loads every document into its context. Runs become slow and expensive, and key details such as an agreed delivery window get lost in the noise. The team restructures the workflow: the agent keeps a short case note, retrieves contract clauses only when needed and discards resolved sub-questions. The context stays small, and the facts that matter stay visible.
Better models alone will not make AI agents reliable and affordable. How context is selected, structured, stored and protected matters just as much. Context Language Models point to a future in which models handle more of this themselves. Until that approach is proven in production, companies benefit most from deliberate context engineering and from keeping their AI stack open to new techniques.
Unitech DACH designs and builds AI agents and agentic workflows, including RAG and Graph RAG on company data. We deliver self-hosted LLMs based on open-source models such as Qwen, Llama and Mistral, run in infrastructure the client controls and adapted via LoRA, QLoRA or full fine-tuning. With n8n, we automate business processes that include LLM and agent steps. Data security comes first: GDPR-compliant processing, data in the EU or wherever the client chooses, role-based access control, and encryption in transit and at rest.
Talk to our experts about your AI use case.
Sources
1. InfoQ — Sergio De Simone: “Context Language Models: Self-Managing Context to Improve Performance and Reduce Compute Costs”, 11 Oct 2026.
2. Shao et al. (University of Washington, MIT, Meta Superintelligence Labs): “Context Language Models”, arXiv:2609.37725, 29 Sep 2026.