AI Daily · 2026-09-28

The day's main thread is agent infrastructure: the Holo4 family launched with 27B dense and 35B-A3B MoE variants that can act through GUI, code, MCP, or API, and its 27B model scores 61.7 on OSWorld 2.0 versus Opus 5.5's 81.8 at what the team frames as much lower per-task cost. LangChain rolled out a broad LangSmith update: Engine v2 adds red-teaming and production-trace diagnostics, Managed Deep Agents gains identity-scoped auth and user memory, and new Trajectories, fine-tuning, and Custom Apps round out the release. A related post pairs TypeSafe's Jev decision model with LangGraph for fast, cheap, deterministic routing and classification tasks, shifting the remaining problem to orchestration. Elsewhere, NVIDIA increased its buyback authorization by $150 billion to $235 billion remaining, and Simon Willison published a year-to-date LLM retrospective plus a Bluesky reply-bot detection tool.

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

Model Release

⭐⭐⭐ [Model Release] Holo4: powering generalist computer-use agents

Hcompany released Holo4, a new series of agentic models in two sizes — 27B dense and 35B-A3B Mixture of Experts — both available on the H Models API, alongside Holotron4 Nano, an updated version of Holotron 3. Holo4 interacts with software through any available interface (GUIs, code, MCP and APIs), with the same model and same invocation across platforms. It was trained via supervised and reinforcement learning on a large set of environments and tasks, including those generated by the company's Agentic Task Factory. On OSWorld 2.0, Holo4 27B scores 61.7% and Holo4 35B-A3B scores 30.9%, against 81.8% cited for Opus 5.5. The team open-sources every trajectory behind its public benchmark scores for step-by-step replay or download on Hugging Face, and says the models compete with frontier models at a much lower cost per task.

Product Update

⭐⭐⭐ [Product Update] New in LangSmith: Engine v2, Managed Deep Agents, Fine-Tuning, and more

At its Interrupt NYC event, LangChain announced a set of LangSmith updates centered on the runtime and observability/evaluation sides of the agent lifecycle. LangSmith Engine v2 adds red teaming, which uses production traces and repos to generate and test hypotheses about issues that have not yet appeared in production, and expands issue detection to cover error rate, latency and cost trends as well as inefficient agent work such as repetitive tool calls and overly long trajectories. For agents on LangSmith Deployment, Engine can now reproduce an issue, test candidate fixes against a broader eval set, and let the user open a PR with one click; the next self-hosted LangSmith release will support BYOK for Engine. LangChain says Engine has analyzed more than 60M traces and diagnosed tens of thousands of issues since launching in May. Managed Deep Agents v0.8 adds identity-scoped auth and user-level memory plus channel and tool management, and the company also introduced Trajectories (a conversational view of traces), LangSmith Fine-Tuning (turning agent sessions into fine-tuning data), and Custom Apps for building and publishing custom LangSmith interfaces such as annotation queues and experiments.

⭐⭐ [Product Update] LangSmith Custom Apps: Build custom interfaces around your agent data

LangChain launched Custom Apps in LangSmith, letting teams build, publish, and run custom UIs on top of LangSmith data directly inside LangSmith. Users can describe the data and visualizations they want in LangSmith Chat and have it generate a working app, or start from provided templates and build against the LangSmith API with a coding agent. Once published, an app becomes a shared interface in the workspace, removing the need to handle hosting, authentication, permissions, and sharing yourself. The feature is available on Plus and Enterprise plans and is aimed at recurring review workflows, such as annotation interfaces where engineers see full traces including tool calls and intermediate steps while subject matter experts see only the request, response, context, and rubric. The post notes that its own human-in-the-loop reviews previously lived across spreadsheets and separate tools, so it built a Custom App for structured LLM evaluation and annotation where reviewers pick a queue, evaluate and edit outputs, and add results to a dataset.

⭐⭐ [Product Update] Bluesky reply bot checker

Simon Willison released a tool called Bluesky reply bot checker, vibe-coded with Opus 5.5, which examines any Bluesky profile for evidence that it is a likely reply bot. It looks for signals such as replies posted within seconds of other posts from the same account, and accounts that never post their own content (or images or links) but instead consistently reply to messages from other, higher-follower users. It also looks for question marks, since the author is especially annoyed by reply bots that make him waste time answering a question no human ever posed. The post notes that automated reply bots are a scourge on Twitter and have started appearing on Bluesky too, and that unlike Twitter, Bluesky still has a freely available and useful API, which makes investigating such accounts less frustrating.

Opinion

⭐⭐ [Opinion] 2026 in LLMs (so far)

Simon Willison published annotated slides and notes from his closing keynote at the WeAreDevelopers World Congress North America in San Jose, walking chronologically through what has happened in LLMs in 2026 so far. He dates the start of his 2026 to November 2025, when Claude Opus 4.5 and GPT-5.1 shipped: incremental improvements over their predecessors, but paired with their coding agent harnesses (Claude Code, Codex) they went from 'often make mistakes' to 'reliable enough to use on a day-to-day basis'. He describes using the 'generate an SVG of a pelican riding a bicycle' prompt to evaluate models and shows the November state of the art, where Claude still couldn't really draw a bicycle and the GPT-5.1 bicycle frame was poor. He also notes the first commit in November to an obscure GitHub repository called 'Warelay', and that he set a New Year's resolution to take on as many new projects as he likes, with 'be more ambitious' as a theme for the year. He mentions sharing predictions for the next one, three and six years on the Oxide and friends podcast with Bryan Cantrill and Adam Leventhal, including that it will become undeniable that LLMs write good code and a predicted 'Challenger disaster' for coding agent security.

Other

⭐⭐ [Other] NVIDIA Announces a $150 Billion Share Repurchase Authorization Increase

NVIDIA announced that its Board of Directors has authorized an additional $150 billion under the company's existing share repurchase program, raising the total remaining authorization to $235 billion.

⭐⭐ [Other] Building Production Agents with Jev and LangGraph

A LangChain blog post explains how to combine TypeSafe AI's newly released decision model Jev with LangGraph to build what it calls 「prod, not god」 production agents. Jev does not generate text; it takes state and a set of questions and returns typed answers with probabilities, described by TypeSafe as a 「system one」 decision model. TypeSafe's benchmarks show Jev running up to 200x faster and 400x cheaper than leading LLMs on narrow decision tasks such as routing and classification. The post says Jev is structured, parallelizable, cheap, and designed to return stable answers across repeated evaluations—in an early Jev-as-a-judge experiment its scores barely moved across 100 runs. The authors argue that once decisions become cheap the challenge shifts to orchestration, which is what LangGraph provides by combining model-driven decisions with deterministic code and addressing context management.