AI Daily · 2026-10-13

Google rolled out Gemini 4 Argon today, opening API access for developers and enterprises and bringing it to Antigravity, the Gemini app, and Search AI Mode; it claims to lead on complex coding, finance, legal, and scientific tasks and ranks first on the Vals Index, with pricing at $2 per million input tokens and $10 per million output tokens. The same model also reaches 87% on the ERQA benchmark with pure reasoning for robotics, and is shown controlling a Spot quadruped and a robot arm. LangChain shipped a wave of Deep Agents updates: reworked MCP support for the new stateless protocol, introduced Deep Agents deploy as an open alternative to Claude Managed Agents, and lifted its Terminal Bench 2.0 score from 52.8 to 66.5 via harness engineering. Other tools include Open SWE, context compression, ScaleDown integration, and LangSmith AI Search; NVIDIA promoted its RTX Spark PC ecosystem and announced new investments in France's startup ecosystem.

North America · First-hand

Google

⭐⭐⭐⭐⭐ [Model Release] Gemini 4 Argon: Built to solve your hardest problems

Google DeepMind released Gemini 4 Argon, a new frontier model available today to developers and enterprises via the API in AI Studio and the Gemini platform, to Google AI Ultra subscribers in Google Antigravity, and to Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search. The company says Argon excels at deep work on complex, real-world multi-step tasks in coding, finance, legal, cybersecurity, and science, and ranks #1 on the Vals Index, which measures performance across critical economic sectors such as finance, coding, legal, and tax work. It launches at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off. Reported use cases include a more than 35% increase in agentic code submissions and nearly a third fewer code rollbacks in Antigravity, a Chrome architectural refactor spanning hundreds of files within Chromium's 40 million-line codebase, a 40% cut in in-memory session prep time and 21% faster cold imports for google/adk-python, up to 2.3x faster molecular structure prediction and 1.8x faster conformer search for the xTB quantum chemistry package, and a 2.85x end-to-end inference speedup for open Gemma models on an A100 via vLLM.

⭐⭐⭐⭐⭐ [Model Release] Gemini 4 Argon: Built to solve your hardest problems

Google announced Gemini 4 Argon, its latest frontier model, available today globally: for developers and enterprises via the API in AI Studio and the Gemini platform, for Google AI Ultra subscribers in Google Antigravity, and for Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search. Google says Argon excels at deep work on complex, real-world multi-step tasks across coding, finance, legal, cybersecurity and science, outperforming other frontier models on benchmarks and ranking #1 on the Vals Index, which measures performance across critical economic sectors. It launches at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off. The post cites internal uses: the Chrome team used Argon to separate core webpage handling from a roughly 6-million-line legacy folder inside Chromium's 40-million-line codebase and complete an architectural refactor across hundreds of files; Google's open-source Agent Development Kit Python codebase saw in-memory session prep time cut by 40% in long sessions and cold imports sped up by 21%; the xTB quantum chemistry package became up to 2.3x faster at predicting molecular structures and 1.8x faster at conformer search. Argon in Antigravity also helped teams increase agentic code submissions by more than 35% while reducing code rollbacks by nearly a third.

⭐⭐⭐⭐ [Model Release] Gemini 4 Argon for Robotics

Google DeepMind released Gemini 4 Argon, which it describes as a step change in physical understanding, code generation and visual-semantic reasoning, applicable across the robotics stack. On the Embodied Reasoning (ERQA) benchmark, which tests understanding and solving of physical, spatial problems in 3D environments, Argon scores 87% using pure reasoning, outperforming other frontier models. As a general-purpose physical reasoning model it spans the robotics lifecycle, and the post shares representative prompts across four areas: robot control (mobile manipulation practiced in MuJoCo, zero-shot robot control, closed-loop steering of video world models), Real2Sim (reconstructing physically calibrated simulation twins from robot camera feeds, egocentric human video or hand-drawn sketches), physical understanding (agentic multimodal hindsight labeling, including dexterity understanding, dense value functions and progress estimation), and research and self-improvement (action improvement via simulation and iterative self-improvement on hardware). Demonstrations include using Argon to control a Boston Dynamics Spot quadruped (12-DoF legs plus a 6-DoF arm) to step on a foot-pedal trash bin and dispose of soda cans, and to control a robotic arm that reorients a fallen mug and places a napkin inside it.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Product Update

⭐⭐⭐ [Product Update] MCP in LangChain: Stateless Protocol, Elicitation, and More!

LangChain has reworked its MCP support to match the new MCP spec released in July. The post cites MCP's official Tier 1 SDKs at close to half a billion downloads a month, and a 98x increase in MCP tool calls from ChatGPT users across 2026, more than doubling in August alone. The new spec introduces a stateless core: a tool call no longer requires opening a session first, the server no longer returns a session ID, so there is no need for sticky routing or a shared session store, and a redeploy no longer kills live sessions. On that basis LangChain now supports tool-list caching, so clients stop re-fetching the catalog every run, and elicitation, letting a tool pause mid-call to ask the caller something, surfaced as a LangGraph interrupt. MCP support has moved into the main package as langchain.mcp, installed via langchain[mcp], with the former MultiServerMCPClient collapsing into a single MCPAdapter class; it is built on FastMCP, which handles transports, auth, connection management and protocol negotiation across both the old and new spec.

⭐⭐⭐ [Product Update] Deep Agents Deploy: an open alternative to Claude Managed Agents

LangChain launched Deep Agents deploy in beta, describing it as the fastest way to deploy a model-agnostic, open-source agent harness in a production-ready way. It is built on Deep Agents, an open-source, model-agnostic agent harness, so users own their memory rather than having it locked into a proprietary harness or a single model. The product bundles multi-tenant orchestration logic and memory, per-session sandboxes, and endpoints for MCP, A2A, human-in-the-loop, and memory into a single command, `deepagents deploy`, where users specify parameters such as model, AGENTS.md, skills, mcp.json, and sandbox. Under the hood it bundles a LangSmith Deployment server with 30+ endpoints, and it uses the MIT-licensed deepagents, AGENTS.md, and Agent Skills open standards while integrating with multiple model and sandbox providers. The post compares it with the recently launched Claude Managed Agents, saying the high-level architecture (harness, agent server, sandboxes) is the same, but the text is cut off at that point.

⭐⭐ [Product Update] All Systems Go: Global PC App Ecosystem Ready for NVIDIA RTX Spark

In a blog post, NVIDIA introduces RTX Spark, describing it as a new beginning for the PC that combines the capabilities of a gaming PC, a creative studio and a personal AI supercomputer into a single superchip. The stated specs are up to 128GB of unified memory and 1 petaflop of AI compute, and the post says it delivers the full stack of NVIDIA RTX technologies. The piece is a product announcement, and its body text is truncated with an ellipsis at the point where the RTX stack is described.

⭐⭐ [Product Update] Open SWE: An Open-Source Framework for Internal Coding Agents

LangChain released Open SWE, an open-source framework for building internal coding agents. The framework draws on patterns observed in self-built systems at Stripe (Minions), Ramp (Inspect), and Coinbase (Cloudbot): isolated cloud sandboxes, curated toolsets, Slack-first invocation, rich context at startup, and subagent orchestration. Built on Deep Agents and LangGraph, Open SWE composes rather than forks, offering an upgrade path as Deep Agents improves and letting org-specific tools, prompts, and workflows be configured rather than modified in core logic. It provides built-in planning via write_todos, file-based context management, and native subagents.

⭐⭐ [Product Update] Context Management for Deep Agents

The LangChain team describes how the Deep Agents SDK handles context management. Deep Agents is LangChain's open source, batteries-included agent harness that lets agents plan, spawn subagents, and use a filesystem to run complex, long-running tasks. To cope with tasks that exceed a model's context window, the SDK applies three compression techniques: offloading large tool results (when a tool response exceeds 20,000 tokens, it is written to the filesystem and replaced with a file path plus a preview of the first 10 lines); offloading large tool inputs (once session context crosses 85% of the model's available window, older write/edit tool calls are truncated and replaced with a pointer to the file on disk); and summarization (when offloading no longer frees enough space, an LLM generates a structured summary covering session intent, artifacts created, and next steps to replace the full conversation history, while the complete original messages are written to the filesystem as a canonical record). These compression steps are triggered at threshold fractions of the model's context window, using LangChain's model profiles to look up the token threshold for a given model.

⭐⭐ [Product Update] Search agent traces by behavior with AI Search in LangSmith

LangSmith introduced AI Search, which lets users query agent traces by behavior in natural language, such as 「did the user get frustrated?」 or 「did PII leak?」. Instead of vector search, a fast decision model reads the actual contents of each trace alongside the query, because embedding a whole trace into one small vector is lossy and can't distinguish an agent that completed a task from one that only said it did. LangSmith uses trajectories to normalize traces, reducing context bloat and speeding up query results. AI Search also composes with existing filters, so a query can be limited to traces that errored or carry a specific tag. The post says that previously, answering such questions required setting up an evaluator ahead of time or going through traces manually.

⭐⭐ [Product Update] Building Long-Running Agents with ScaleDown and Deep Agents

LangChain released the langchain-scaledown integration and added support for ScaleDown's constrained decoding models (CDMs) in the LangSmith LLM Gateway. A CDM is trained for a single task and only returns answers that task allows, such as a label from a list or spans of the input, making it much smaller, faster, and cheaper than a general-purpose LLM; ScaleDown reports its models are 200x faster and 467x cheaper than frontier models on these narrow tasks. The integration ships with three middlewares for LangChain and Deep Agents: one summarizes old history, one compresses large tool results, and one extracts structured data from the conversation. ScaleDown offers four single-task models — Summarize, Compress, Extract, and Classify — each taking up to 1M tokens of input, with Extract, Summarize, and Classify also accepting images and scanned PDFs. On ScaleDown's own benchmarks against the average of GPT-5.5, GPT-5.4 Mini, and GPT-5.4 Nano, Summarize is 41x cheaper and 2.7x faster, and Compress is 138x cheaper and 1.8x faster, with both scoring higher on accuracy.

⭐⭐ [Product Update] 3.4.1

DSPy has released version 3.4.1. The release makes GEPA's structured InstructionProposer the default instruction proposer: it calls the reflection LM through dspy.Predict with a JSONAdapter, so dspy.History inputs, tool trajectories and dspy.Type values such as dspy.Image reach the reflection LM as structured content instead of text placeholders, and it can be configured via skills, additional_instructions, base_instructions, max_chars, truncate_history_outputs and adapter, with MultiModalInstructionProposer still available. dspy.RLM can now import dspy inside its REPL and build sub-agents (dspy.Predict, dspy.ChainOfThought, dspy.ReActV2 and nested dspy.RLM); sub-agents are built and run on the host, only JSON crosses the interpreter boundary, and max_llm_calls also counts sub-agent LM calls. The version adds Python 3.15 support, fixes a 3.4.0 import crash, and includes streaming, adapter and interpreter fixes. The notes also list compatibility changes: the reflection request changes shape, the reflection LM must be a dspy.LM that can return JSON, ReActV2 signatures can no longer declare inputs named history or tools, and programs saved on 3.4.0 with a null Score level no longer load; legacy forward()/aforward() custom LMs, legacy-engine wrappers, complete_legacy() and OpenAI-style messages= calls are deprecated for removal in 3.5.

Research

⭐⭐⭐ [Research] Improving Deep Agents with harness engineering

LangChain described how its coding agent deepagents-cli improved on Terminal Bench 2.0 from 52.8 to 66.5 (+13.7 points), moving from outside the Top 30 into the Top 5, with the model held fixed at gpt-5.2-codex — only the harness changed. The team compressed the tuning space to three knobs: system prompt, tools, and middleware. They packaged trace analysis as an Agent Skill that fetches experiment traces from LangSmith, spawns parallel error-analysis agents, and has the main agent synthesize findings and suggestions into targeted harness changes. The post notes the most common failure pattern was the agent writing a solution, re-reading its own code, and stopping once it looked fine, so the improvements focus on pushing agents into a build-and-self-verify loop via system-prompt guidance on planning, discovery, and verification.

Opinion

⭐⭐ [Opinion] Quoting Geoffrey Huntley

Simon Willison quotes Geoffrey Huntley on getting coding agents to use a designated diagnostics library. Huntley notes that if the agent isn't using the library, he could introspect and update AGENTS.md, but a gap remains and will drift. Instead, he writes a pre-commit hook that checks the git diff for direct calls to observability modules that bypass the diagnostics module; when found, it fails the commit and injects a prompt telling the agent to use the diagnostics library. He says the result immediately became ten out of ten. The quote comes from a piece titled “nobody should type code by hand anymore. here's what's left.”

Other

⭐⭐ [Other] NVIDIA Announces New Investments in France’s Startup Ecosystem

NVIDIA announced new investments in France's startup ecosystem. The post says companies in France are deploying agentic computing across industries including space, financial services, the public sector and scientific research, showing the country's ecosystem has moved from ambition to application. It notes that the February 2025 AI Action Summit and the 2025 and 2026 Choose France Summits laid the groundwork and anchored major compute commitments.

⭐⭐ [Other] How We Made Our Docs Test Themselves

A LangChain engineering blog post explains how the team made the code samples in its docs testable. The author notes that inline samples go stale as APIs and dependencies change, and that the manual migration—extracting code into standalone files, adding setup and teardown, adding snippet markup, and wiring it into CI—carries enough upfront cost that the project often never happens. Their approach was to delegate the work to the Deep Agents CLI using Skills: the migration steps (move code into src/code-samples/{product}, make samples complete and runnable, lint, tag snippets with :snippet-start:/:snippet-end:, run the samples, and generate snippets back into the docs) are written as reusable instructions. A GitHub Action runs the tests regularly and files tickets when a test fails; the SKILL.md lives at .agents/skills/docs-code-samples/ in their docs repo, and anyone can open the Deep Agents CLI there and ask the agent, for example, to migrate the inline code in streaming.mdx to testable samples.