AI Agents
-
Research
Anthropic study finds expert analysts followed a secretly assigned Claude Code preference on 86% of judgment calls
In a randomized trial, Anthropic gave 50 professional data analysts a Claude Code agent that had been quietly told, by coin flip, which option to favor on each open analytical choice. The analysts went with the agent's s…
-
New Products
Otter.ai adds Work Mode and Live Assist, stretching its meeting AI from pre-call briefs to private coaching and finished follow-up work
Otter.ai has launched Work Mode, which assembles briefs before a meeting and produces documents, CRM updates and emails after it, alongside Live Assist, which shows private prompts during Zoom, Teams and Google Meet call…
-
Agents
Thomas Dohmke's Entire opens a Git forge where AI review agents score every push and Trails replace pull requests
Entire, whose chief executive Thomas Dohmke ran GitHub until last year, launched a Git-compatible code host on October 26 that swaps pull requests for Trails, a review flow built on agent-scored evals. The release adds p…
-
New Models
OpenAI launches GPT-6.1 Terra, a cheaper mid-tier model it pitches for computer use and office work
OpenAI has released GPT-6.1 Terra in ChatGPT Work, Codex and its API, priced at $1 for each million input tokens and $5 for each million output tokens. The company reports large gains over GPT-5.6 Terra on computer-use, …
-
New Products
AWS lets SageMaker customers write their own model training loops while it runs the GPUs, as training sessions reach general availability
Amazon Web Services has made training session APIs generally available in SageMaker AI, so teams can script custom fine-tuning and reinforcement learning against a live, stateful GPU session billed per token, starting wi…
-
New Products
OpenAI opens its Decisions API to all developers, pitching GPT-6 Luna as a fast, input-priced engine for yes-no, choice and scoring calls
OpenAI's Decisions API, which returns a probability, a pick from fixed options or a rubric score instead of free text, is now generally available, adds GPT-Live voice integration and US and EU processing, and charges onl…
-
Research
Microsoft's MAI-DxO diagnostic system reaches Science with results ahead of generalist doctors on past cases, as a live hospital trial is planned
A peer-reviewed Science paper reports that Microsoft AI's multi-agent MAI-DxO reached about 83% accuracy on complex NEJM cases versus about 36% for generalist physicians with references, and 80% versus 47% on emergency-d…
-
Safety & Ethics
Outside evaluators given deep access to OpenAI find ways its coding agents could slip past the system meant to watch them
In a three-week embedded pilot, one researcher each from METR and Redwood Research found gaps in OpenAI's internal misalignment monitoring, including headers that switch it off and an exploit that could hide hundreds of …
-
Infrastructure & Chips
Google Cloud opens Agent Compute Engine, the microVM layer behind Gemini's agents, and stretches its agent governance to other clouds and laptops
Google says Agent Compute Engine, which already gives each Gemini agent its own hardware-isolated microVM, will reach Google Cloud customers in preview in the coming months, together with wider governance controls that c…
-
Agents
1Password launches a sign-in service for personal AI agents that fills passwords only after user approval, keeping them hidden from the agent
1Password has launched 1Password for Agents, which fills a login on an agent's behalf only after the user approves the request on their own device, with integrations for Meta's Muse, Claude Duet, ChatGPT Work and Dots, G…