Agents
94 articles RSS
Researchers Coin 'Coherence Debt' to Explain Why Coding Agents Fabricate Fixes Instead of Asking When Facts Go Missing
A study spanning seven AI models and five coding-agent harnesses finds they fail identically when a required repository fact is unavailable, and fabricate rather than stop.
New Study Finds AI-Generated Bug-Fix Patches Far More Bloated Than Developers', Proposes RECAP to Slim Them Down
A new study finds AI-written bug-fix patches are far larger and more complex than developers' own fixes even when tests pass, and proposes a lightweight adapter called RECAP to trim the bloat.
DeepSeek Open-Sources Harness Agent Framework Under MIT License, Ships V4-Pro With Native Responses API Support
DeepSeek released its Harness agent runtime as MIT-licensed open source and rolled out DeepSeek-V4-Pro with native OpenAI Responses API support.
OpenAI, Amazon, Microsoft, Cursor and Vercel Launch Agent Plugins 1.0 as Google Joins the Open Standard on Day One
Six companies agreed on a shared packaging format for AI agent skills and MCP servers; Anthropic, which created the underlying protocols, is not part of the effort.
Bank of America Adds Generative AI to EricaAssist, Cutting Customer Service Call Times by Nearly a Minute
Bank of America has enhanced EricaAssist, the AI agent used by more than 18,000 customer service representatives, with generative AI that delivers guidance in under three seconds.
Cloudflare Launches Kitesurf, a Browser Built for AI Agents That Runs Entirely on Its Workers Platform
Cloudflare's new Kitesurf browser skips Chromium entirely, running in V8 isolates on Workers to cut CPU and memory use for AI agent tasks.
Anthropic Launches Claude Science, a Research Workbench for Drug Discovery, and Starts Its Own Neglected-Disease Drug Program
Anthropic released Claude Science, an AI workbench for scientists with more than 60 curated skills, in beta to paid Claude subscribers, and said it has started an internal program to develop treatments for neglected diseases.
WorkBench Revisited Finds Top AI Agent Now Completes 89% of Office Tasks, Up From 43% in 2024, With Safety Improving Alongside It
A two-year follow-up to the WorkBench benchmark finds the best workplace agent jumped from 43% to 89% task completion while unintended harmful actions fell from 26% to 2.5%.
Model Context Protocol Locks Its Largest-Ever Spec Revision, Rebuilding the Core as a Stateless Protocol Ahead of a July Release
The 2026-07-28 MCP release candidate makes the protocol core stateless, formalizes an Extensions framework, and deprecates Roots, Sampling, and Logging. The RC locked May 21 for a July 28 final.
Mistral Renames Le Chat to Vibe, Folding Chat, Office Automation, and Cloud Coding Into One Agent
Mistral rebranded its Le Chat assistant as Vibe, a single agent spanning Work Mode for office tasks and Code Mode for remote coding.
OpenAI and Molecule.one Report a Near-Autonomous AI Chemist That Improved a Stubborn Drug-Making Reaction Across 10,080 Experiments
GPT-5.4 paired with Molecule.one's Maria platform proposed, ran, and analyzed a 10,080-reaction campaign that lifted yields of a hard sulfonamide coupling using the additive TEMPO.
DeepMind's AlphaProof Nexus Solves 9 Open Erdős Problems by Pairing Gemini With the Lean Proof Checker
A DeepMind preprint reports an LLM-and-Lean agent that autonomously solved 9 of 353 open Erdős problems and proved 44 of 492 OEIS conjectures for a few hundred dollars each.