로컬 실행과 에이전트 검증, GPT-6 Astra와 바이트 모델의 부상

Aravind Srinivas

@AravSrinivas

1일 전

We’re expanding our work with @nvidia to bring fully local AI to Microsoft Windows PCs with RTX GPUs. Unmetered local intelligence on every Windows PC running on NVIDIA hardware and Perplexity harness. Enjoy!

Powerful AI agents are becoming easier to run right on your PC. Together with @Perplexity_AI , we’re making local AI more private, accessible and useful on @Windows NVIDIA RTX PCs.

Introducing Cline Desktop - a native interface for working with open weights models. Use with ClinePass and all our free models like DeepSeek-V4.1-Flash, Musespark-1.3, or BYOK with any provider!

I think the IDE is officially dead. The Cline desktop app is here. Open-source to use it with open-weight models. Notice how we are moving away from the terminal. Every agentic coding tool started with a TUI, but we need better affordances that only a desktop app can provide. Terminal → Desktop app → [...]

On building an agent harness from scratch. Got so many questions about where to get started. My short guide (feed it to your agent): If you really want to learn harnesses well, it's worth building one from scratch using a programming language (TypeScript or Python) of your choice. When I got started, I implemented my first harness using ReAct from Google: https:// academy.dair.ai/papers/react-s ynergizing-reasoning-and-acting-in-language-models-2210.03629 … At the time, I built this from scratch, but you can easily prompt your agent to consume the paper and produce a minimal implementation you can inspect and understand. You want to target having three parts: - an LLM module for all things inference, and it should ideally support several models. I used OpenRouter when I got started. This can include the system prompt, but you can also separate it out if you plan to explore context-engineering ideas more deeply. - a tools module (I recommend building them as MCP tools for interoperability, but you can design functions from scratch if you have experience). - an agent loop that encapsulates the tools and LLM. ReAct is one of the more basic loops you can implement. Primarily, aim to understand the main components and how they work with each other. Pro tips: - try to keep your system prompt minimal and experiment with different models; a mini version of all frontier lab models should be good enough to get you started. - look at the code and log things as you experiment with different tasks. You want to log inputs/outputs to the loop, inputs/outputs from LLMs, and inputs/outputs from tool calls as a starting point. Set up a simple set of diverse tasks to test your agent loop on. So with every change, you can run the tasks and inspect the results manually. Once you have a good grasp of this, you can easily add other things like skills, memory, etc., once you have a good idea of how to tune them. It helps to keep things modular if you are planning for this. I would recommend playing with memory, skill, and subagent as good next steps. If you don't want to build the components or want to start building a more serious agent harness, I recommend using the Pi SDK or LangChain harness tools. I am also going to release something soon to help with this. Let me know if you have questions. I am planning a longer write-up on this, but this should be enough to give you something to experiment with.

Caspar

@caspar_br

1일 전

good overview on the managed agents landscape by @Sam_Witteveen we've seen many major ai companies ship a managed agent product over the past few months. this space will only continue to grow exponentially langchain first built a managed agent product two years ago - langsmith deployment - a self-hostable runtime for agent workloads. LSD has evolved with our open source ecosystem and is used today for hosting anything from simple langgraph workflows to complex apps built on deepagents harness this space has come a long way in two years. managed deployment was the first step. the next is to give devs a complete foundation for their agents, so they don't have to assemble it themselves. that's why we built managed deep agents: you bring the the agent's intelligence (its instructions, skills, tools, model) and we provide a managed runtime (LSD agent server, session management, credentials, tracing) out of all the options listed below, MDA is the only one built on an open source harness. if you're building agents, see how much time you could save: http:// langch.in/mda

OpenAI Developers

@OpenAIDevs

1일 전

GPT-6 Astra in Codex is helping test code end to end at @perplexity_ai . @randomjohnnyh uses it to build test harnesses and mock third-party API responses, so he can check how the pieces work together.

GPT-6 Astra tops the ARC-AGI-3 leaderboard while cutting token costs. Here is what matters for developers: Tied with Claude Fable 5.1 on Artificial Analysis’ Intelligence Index Asynchronous tool calls for parallel execution Retained reasoning memory across API calls Efficient context management is essential to scaling agents. Read our full analysis in The Batch: https:// hubs.la/Q04xp3NY0  #DeepLearningAI #AIAgents #MachineLearning

Great paper from Amazon. In discusses when not to trust LLM judges for agent evaluation. (bookmark it) A common way to compare task agents is to have an LLM user simulator talk to each one and an LLM judge score the transcript. This paper from Amazon shows that gate fails in two specific ways. 1. Satisfaction does not track success. 57.5% of conversations the raters marked satisfied had failed the customer's task. 2. Close calls go wrong. The ranking holds across agents of very different ability, but among near-equal agents the gate picks the lower-reward one on 31% of pairs, compared with under 1% for pairs far apart. The study covers 25 agents from six providers on tau2-bench and SimulatorArena. Judges also favored agents from their own model family. The fix is cheap. A judge-free completion bit catches truncation regressions, and the judge is trusted only after calibration against a verifiable reward. Paper: https:// academy.dair.ai/papers/gauge-w hen-not-to-trust-llm-as-a-judge-in-user-simulated-evaluation-of-task-orie-2609.12191 …

If you are tracking agent swarms research, this one is worth reading. Between 24 May and 2 July 2026, autonomous agents running inside a timed research evaluation wrote to a third party's public wiki. OpenAI acknowledged the incident. This paper reconstructs what happened from the wiki's archived history, 14,591 revisions across 4,579 pages. It identifies 907 agent cohorts and estimates about 876 episodes. Coordination formats converged within a day. Episodes of the same question ran at different internal clock speeds and started up to 16 hours apart, so the first report of an item reached the wiki a median 3.4 hours before a later cohort arrived. Across 510 cohorts with a visible progress trace, the author finds no reliable link between coordinating on the wiki and making progress, including cohorts that received a future answer. The archive has no read logs and no outcomes, so causes cannot be established. The paper's recommendation for anyone running agent evaluations is to log reads and outcomes. Paper: https:// academy.dair.ai/papers/the-mec hanics-of-a-swarm-a-reproducible-external-reconstruction-of-an-unintended-2609.12748 …

Banger paper from Meta. This work shows that byte-level models start out behind token models and then pass them as compute grows. They show this for distilled 1B models trained on up to 1 trillion bytes. To distill a byte student from a token teacher, they convert the teacher's token logits into byte logits, either approximately (Marginalize-It) or exactly (End-Of-Token). Token models lead at low compute but plateau. Byte models reach a higher ceiling, and the fitted scaling laws predict the End-Of-Token model ends up to 4% ahead of the distilled token model. The byte models also match the distilled token model with one-sixth of the training data, and a 256-entry vocabulary cuts teacher-logit storage to about a fifth. Paper: https:// arxiv.org/abs/2609.12303 Chat with Paper: https:// academy.dair.ai/papers/breakin g-the-token-ceiling-distilling-smaller-stronger-byte-models-2609.12303 …

Runway Labs

@runwayml_labs

1일 전

Real-time video models are unlocking new capabilities. A first experiment: a drawing app that animates your sketches as you draw. Under the hood, a video model generates each frame live, turning every idea into motion instantly.

Don’t pace building and owning your own AI models and systems as an enterprise, and there will be no doomsday for you

China is not chasing AGI or the Singularity. They never write “Singularity is our national horizon.” They do write that AI evolution will be vast, and everything so far is a drop in the ocean. A 5-level map to true RSI: Recursive self-improvement just got a 5-level checklist. A 33-author Chinese survey already named the destination: “The Last AI Built by Humans.” - level 1 AI executes the patch - level 2 AI picks the strategy - level 3 AI gathers its own experience - level 4 AI adapts from the real world - level 5 AI rewrites the improver itself L5 is the singularity-shaped claim. You stop building models. You build the process that builds the next process. Most systems today: L2–L3. L5: early evidence, not ignition. RSI is not “the model got better at coding.” It is this: - experience to persistent change - that change to better future improvements - those improvements to inherit into the next round If the loop only makes outputs better, that is refinement. If the loop makes the improver better, that is RSI. They introduce HCI Headroom-Closed Index. Point: current LLMs are not infinitely improvable by reading more human text. Progress is uneven. Interactive work (coding, tools, agents) still has a lot of closed headroom. Scaling a static model is not a closed improvement loop. So they stop asking “how big is the model” and start asking “who owns the next improvement.”