A recent finding shows models describing, fixing, and even writing the weights of smaller networks from scratch, without running training. The technique, dubbed weight space programming, is still limited to games and RL policies, but it points to a real shortcut in the AI research cycle.
Legora's CTO, the Swedish legaltech, explains in an interview with the Latent Space podcast how the team debugs and evaluates an agent that handles confidential contracts without access to a single real usage session.
Jepsen creator Kyle Kingsbury raised a warning: LLMs have made it easier to fake maturity in software projects. For those who choose dependencies day to day, this changes what's worth checking before trusting.
In an interview with the Latent Space podcast, two Anthropic engineers detail the architecture that isolates the model's reasoning from task execution, the piece that underpins Claude's long-running agents.
New version of the datasette-llm plugin, released on October 31, 2026, now uses the default model already configured in LLM, eliminating a manual step for those who want to test Datasette Agent locally.
On October 29, Google and CopilotKit published a guide showing how agents built with LangChain, Google ADK, AWS Strands, and Mastra can draw reactive interfaces, not just text, using the open A2UI standard.
On the Latent Space podcast, Arca's lead AI engineer explains how the fintech sustains financial tasks that take up to ten days without letting the agent decide on its own. The architecture combines persistent per-customer memory, a custom harness, and deterministic guardrails beyond the model's reach.
Anthropic's hiring manager revealed, this October 2026, that he stopped giving candidates coding challenges because Claude already solves these tests better than any human. For those hiring developers, the question that remains is: what to measure instead?
The Model Context Protocol's Authorization specification, published in the June 2025 revision, defines how MCP servers exposed over HTTP must validate OAuth 2.1 tokens. Anyone who deployed a server without this layer risks falling out of compliance and breaking when stricter clients start requiring the full flow.
In the game Spawn, autonomous AI agents ship code to production every hour and decide on their own what's a bug, what's a feature, and what's not worth fixing.
A walkthrough published by Google on October 28, 2026 shows the Agents CLI's evaluation loop: a small dataset, a numeric score for every change, and an automatic alert when the agent regresses.
Simon Willison tested Claude Opus 5.5 on four complete games and noticed something new: models finally understand what makes a game fun, not just visually convincing. That changes the math for anyone who wants to prototype games fast.
At GitHub Universe 2026, the company announced that merged pull requests reached 220 million per month and commits topped 7.38 billion a month in less than a year. The numbers are real, but what they mean for whoever reviews that code is another story.
In an episode of the Latent Space podcast recorded in October 2026, Anthropic executives detail how the Claude Platform separates an agent's reasoning from code execution, a pattern that reduces latency and isolates credentials.
A study from the Allen Institute for AI tested six models as
A walkthrough documented by a Thomson Reuters engineer shows how to set up a coding agent that runs entirely on a laptop, with no cloud, using OpenCode, Ollama, and the open model Qwen3.5.
Ollama has supported structured function calling for a while, but the wave of models released between October 14 and 24, 2026 made this feature much more usable for building truly local agents.
Simon Willison's review shows that webframe, by Jeff Lindsay, runs pages like a script without a bundled browser, but runs into CSP, a zeroed-out viewport, and stuck page loads.
The 2025-06-18 revision of the Model Context Protocol, published on June 18, 2025, removes batch call sending, requires explicit version negotiation, and adds elicitation and structured output. Anyone maintaining an MCP server needs to review their code.
The 2025-06-18 revision of the Model Context Protocol, published in June 2025, formalized elicitation, structured output, and resource links, but it also removed JSON-RPC batching. Anyone already running an MCP server needs to review transport and authentication before updating the SDK.
Models solved math problems that had been open for decades using deep search and automatic verification. The same pattern explains why coding agents work well with tests but stall when there's no clear right-or-wrong criterion.
Simon Willison used a single prompt to get Claude Opus 5.5 to generate an artifact that paints photos using Sobel edge detection, running entirely in the browser without any libraries.
Researchers at Calif and Wiz independently discovered two critical flaws in the PHP core using GPT-5.6 Sol and Gemini 4 Argon; the fix took two months of coordination before shipping in October 2026.
