Search: “context”

Context compression is a huge bottleneck for long-running agents. This work finds that context compression hurts long-horizon agents at a few specific points. They propose PAIR, which replays the agen

a big part of agent harnesses is how they interact with context memory is just context its therefor impossible to separate harness from memory - as @sarahwooders says, 'memory isn't a plugin (it's a h

// Reusable Context Engineering // Context bloat quietly kills long-horizon runs, but you can fix it from the outside without fine-tuning the underlying agent. (bookmark this) Context management is us

The Harness is a Context Manager on Behalf of the Model What happens when the context window fills up and who decides? This decision is external to the model - The Harness designer must have some opin

Great paper for the weekend ahead. Meta presents Context Language Models, which can natively manage their own context. Banger paper from Meta and colleagues. (bookmark it) They discuss benefits of giv

Your agentic workflows are only as good as the context you feed them. In finance, that context is stuck inside your messiest documents: dense tables, footnoted adjustments, and the details buried in t

We launched Context Hub as a way to manage skills, AGENTS.md files, and other context files an agent might need You can easily use it as a virtual filesystem in deepagents See this video for more info

Model. Harness. Context. The 3 main components of agents. As you build more agents, context increasingly lives AGENTS.md, skills, policies, examples, + generated research files. Context needs its own

Everyone talks about 1M context. The harder part is making 1M context actually usable. Serving MiniMax M3 required optimizing for long-context, multimodal, and agentic workloads simultaneously. Excite

We've productionized query-aware compression for faster, cleaner, more-accurate search. Better context is better than more context. Our system cuts context tokens up to 70% while improving answer qual

Cursor 3.3 introduces a context usage breakdown feature that allows users to view detailed statistics about how their AI agent consumes context tokens. This diagnostic tool helps identify context-rela

I was unable to retrieve the specific X (Twitter) post at the URL provided, as the content is not accessible through search results. The tweet's text snippet ('memory is context, and context is wha...

Nice paper discussing context trimming for agents. This is a hot topic at the moment, so it might be worth your time. Context trimming for agents is usually judged by how many tokens it removes. This

// The Efficiency Frontier in LLMs // (bookmark this one) How much are you overpaying for context you do not need? It turns out that context costs dominate production LLM bills, and the right strategy

Jörg Schad is the VP of Engineering at @pinecone. He joins @kbal11 to discuss moving beyond RAG with precomputed context, context artifacts, semantic layers, and AI retrieval. @joerg_schad https://sof

Introducing Qwen-Audio-3.0-ASR-Flash: More context-aware. Stronger domain-term recognition. 🚀Our latest ASR model upgrades: • Context consistency • Domain-term recognition • Custom hotwords • Speech p

you can now build RLMs with Deep Agents! the more context agents accumulate, the worse they perform, a phenomenon called context rot. RLMs, proposed by @a1zhang from MIT, help: instead of working mode

We trained language models that compress massive contexts into tiny latent representations. Latent Context Language Models (LCLMs) outperform existing KV cache compression methods on the latency/accur

Cursor can now show your agent's context usage as an interactive report in a canvas. The context explorer breaks down where tokens go across the system prompt, tool definitions, rules, skills, and mor

100% agree on the Context Hub. Developers constantly tweak their approaches to manage context in their prompts with each new model and tool suite release. I would even say that the core problem we fac

looks like new Pareto frontiers across everything: - Context: 400K context in Codex and a 1M in API - API Pricing: 5/m input and 30/m output tokens. - Codex improved its own inference speed 20% lol -

You can also set your autocompact threshold yourself and effectively lower your context window if you'd prefer. For example, 400k context is a good compromise: CLAUDE_CODE_AUTO_COMPACT_WINDOW=400000 c

deepagents now supports structured output for subagents! an under appreciated piece of context engineering is figuring out what context is passed back from subagents to the main agent. now you can def

Interesting paper from Tencent. Tencent trains an agent to manage its own working context, and assigns credit at the level of individual context edits. Long-horizon tasks force a model to retrieve, in

Great overview of context compression in LLM Agents And one interesting, unexpected finding. If your compaction policy is tuned to cut tokens, it may be making your agent run slower. Great study from

This post completely misses the point of @sarahwooders 's original article. The whole point is that memory == context engineering, so it can't be a 'thin conductor'. Saying 'memory is just markdown' t

Bullish on this trend of making post-training more accessible. A new post-training era is upon us. If you work on agentic RL, long-context tasks (a big focus today) are expensive, inefficient, and don

Four small architecture decisions can cost up to 47% of a model's long-context performance. New research from Ai2, Carnegie Mellon, and the University of Washington isolates them. Normalization, GQA,

New research from Meta and CMU. This one is on agentic context management for long horizon tasks. (bookmark it) Production agents accumulate context every turn. The usual fix compresses on a token thr

We scaled a robot model natively to 8,000 timesteps of context, 5 minutes worth of muscle memory, with constant inference cost. Robot policies used to live their lives a few frames at a time (< 0.1 se

Excited to share our new paper: RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably LLMs often fail on inputs well within their advertised context lengths. We show that these fa

Banger paper from Meta and colleagues. (bookmark it) They discuss benefits of giving your agent write access to its own context. In other words, they investigate how effective it is to allow a languag

// Agentic Context Management // Great read for the weekend. (bookmark it) Production agents fail less on reasoning and more on what sits in their context. Conversation history, big prompts, huge tool

// The Efficiency Frontier // Cool paper on context management. As agents reuse the same documents and histories across many turns, the cheapest context strategy is not fixed. This work describes a pr

Most illuminating graph I have seen for definition of harness. Memory and context are deeply coupled with harness. In my humble opinion, how they are injected and managed are one of most important pie

Banger paper from NVIDIA on long running agents. A model can accept 128K tokens of context and still make more mistakes the longer it works through a task. If your agent loses its place partway throug

At a 1M-token context length, QSA’s attention kernel is up to 7.6× faster in prefill and 4.9× faster in decode. With a 90% prefix-cache hit rate, Qwen3.8-Flash-Next delivers 8.6× the prefill throughpu

I am bullish on centralized context in agentic engineering to steer what AI produces, in a way that reduces ai slop potential across the sdlc. To bring in engineering context, I need integrations conn

This 1 hour workshop on context engineering will help you build agents that hold long conversations and cut your token cost by up to 90%. Last month we gave it live at the AI Engineer World's Fair 202

I edited the intro because I realized I buried the lede originally- The 1M context window is a double-edged sword. It allows Claude to do more complex tasks but it can also leads to more context pollu

Harness, Memory, Context Fragments, & the Bitter Lesson this is a work in progress mental dump on interesting intersections between how we use and design a harness, implications for memory being accum

Introducing 'Scaling In-Context Imitation Learning' (SAIL) to be presented at #IROS2026. This work is a collaboration between Sakana AI and the University of Tokyo. Blog: https://pub.sakana.ai/sail Wh

// ContextLeak in AI Agents // The whole attack surface here is a tool name and a tool description. Stealing an LLM agent's runtime context, meaning the user prompt, the execution trajectory and the t

'Agents = model + harness + context' is a much better mental model than treating the LLM as the entire product. The model provides capability, but the harness determines tools, permissions and behavio

Does your context compactor actually save anything? If you built or tuned one, this one is worth your time. (bookmark it) Task completion is the standard way to validate compression, and it hides the

Qwen 3.8-Max is live on Modal, with the full 1M context window and a custom DFlash speculator under the hood. Love seeing our launch partner Modal go all in from Day 0!⚡️ 2.4T parameters. 1M context.