13 stories
Anthropic's Project Deal pilot had 69 employees use AI agents as buyers and sellers in a real-goods marketplace, completing 186 deals totalling more than $4,000 — and found that users with better models got measurably better outcomes without noticing the gap.
METR and Epoch AI have jointly published preliminary results from MirrorCode, a project examining whether current AI systems can complete software engineering tasks spanning days or weeks of human effort.
Databricks moved Lakeflow Designer from early access to public preview, offering analysts and domain experts a drag-and-drop canvas and natural language interface for data preparation that generates production-ready Python code.
Adept has begun rolling out access to Adept Experiments, starting with a plain-language workflow builder for the web powered by ACT-2, a model fine-tuned from the Fuyu family for UI understanding and action-taking.
Adept has published a capability preview of ACT-1, a transformer model hooked up to a Chrome extension that lets it observe a browser viewport and perform click, type, and scroll actions in response to natural language commands.
Adept has published details of AWL, a JavaScript ES6 subset that lets users compose browser-based agent workflows using a mix of prescriptive function calls and natural language, with the company framing it as the authoring layer for production enterprise automation.
Publication held pending source verification.
The Series B, led by Ribbit Capital with Sequoia and others participating, comes nine months after launch. The company says it has conducted over one million interviews and grown annualized revenue 15x to eight figures, according to VentureBeat.
Researchers Haebin Seong, Li Yin, and Haoran Zhang present a framework in which a Harness Evolution Loop optimizes a worker agent's prompts, tools, and orchestration for a single task, while a Meta-Evolution Loop learns an evolution protocol the authors say enables rapid harness convergence on any new task without human engineering.
Cowork, available as a research preview for Claude Max subscribers on macOS, gives non-developers access to the same agentic loop that powers Claude Code — read, edit, and create files in a designated folder.
The new Slackbot, now generally available to Business+ and Enterprise+ customers, moves from algorithmic notification tool to an LLM-backed agent capable of searching enterprise data, drafting documents, and taking action.
Import AI 453 covers a benchmark from METR and Epoch showing AI autonomously reimplementing complex software, a Google DeepMind taxonomy of agent attack vectors, and the Windfall Policy Atlas for navigating transformative AI policy.
Researchers introduce COSPLAY, a framework in which an LLM decision agent retrieves skills from a learnable skill bank while a separate pipeline continuously extracts and refines those skills from unlabeled rollouts. An 8B model achieves over 25 percent average reward improvement against frontier LLM baselines on single-player benchmarks.