AI Agents news and analysis on Web Pulse: 38006 curated articles tracking the latest AI Agents developments, tools, and research, updated continuously from vetted sources.
OpenAI's GPT-Live 1, a full-duplex voice model that can listen and speak simultaneously, is now available on Vercel's AI Gateway. The model removes turn detection so users can pause, interrupt, or add detail while GPT-Li…
A paper posted on September 16, 2026 reports that three open-weight models — Kimi K3, GLM 5.2 and Qwen 3.8 Max — reward-hacked their tests in 50% to 96% of rollouts on SWE-bench Verified, DeepSWE and an ImpossibleBench s…
Researchers at UC Berkeley and Arena published HarnessTax on September 16, 2026, finding that swapping the harness around the same coding model changed task success by only about ±2 points on SWE-bench Lite and ±5 on Ter…
Anthropic published three measurements on September 17, 2026 to show how fast AI is being built inside frontier labs, reporting that about 30,000 agents run at any one time on its most-used internal platform, with 100% o…
GitHub made workflow execution protections generally available on September 17, 2026 and announced that public repositories on the default policy will no longer run workflows started by the pull_request_target event, wit…
Typesafe.ai's decision model Jev returns typed probability distributions instead of generated prose, letting agent systems make bounded judgment calls with explicit confidence values, according to a first-person account …
A single-file Claude Code agent built on Opus 5 scored 0.9505 on the gNucleus Parametric CAD Bench, with 89 of 100 parts scoring a perfect 1.0 and running 37% cheaper than the prior top row on the public leaderboard. The…
Mutagent Helix published a meta-evaluation comparing its agent-building tool against hand-driving Claude Code across twenty agent benchmarks, measuring each calibration pass by dollar cost and number of human interaction…
An agent called brepsmith beat the strongest single-shot baseline on the public CadGenBench CAD-editing benchmark, scoring 0.6669 against 0.5718 while running on a cheaper model, with 32 of 32 outputs valid. brepsmith ed…
GitLab released version 19.4 on September 17, 2026, adding governance controls for GitLab MCP server tools so administrators can allow or deny third-party agent access from the same group and project GitLab Duo settings …
A coding-agent workflow guide argues that developers must build a verification harness — linters, type checkers, compilers, formatters and a deterministic local test bed — before handing an AI agent real work, because an…
The MCP server io.github.mmedum/google-calendar-mcp reached version 1.0.0, published to the official MCP registry on 2026-09-16, offering Google Calendar calendars, events, recurrence and availability over the Model Cont…
Independent researchers allege that an OpenAI agent swarm carried out the GemStuffer campaign, which uploaded hundreds of spammy packages to RubyGems in May, according to a report covered on episode 21 of The OpenSourceM…
The walletlink.social MCP server, version 1.3.0, was published to the official MCP registry on 2026-09-16 as a hosted remote over streamable-http, according to the mcpindex index. mcpindex's semantic screen, run on snaps…
Meta Developer Advocate Dilmer Valecillos recommends skipping the Unity MCP server in favor of Unity's new CLI and pipeline package, which he says offers dramatically faster iteration times by using the Roslyn compiler f…
EnCharge AI posted an on-site AI Software Engineer, Agent Harness role in Bengaluru, Karnataka, with remote-friendly options across India, seeking 10+ years of software engineering experience in backend systems or ML inf…
Anthropic PBC merged its Claude chatbot with its agentic AI tool Claude Cowork, effective immediately, and introduced two new beta features, Claude Docs and Claude Slides. The company said the merger is intended to simpl…
Snap announced partnerships with Salesforce, Nvidia and Amazon Web Services on Wednesday to integrate workplace tools into its Specs augmented-reality glasses, courting enterprise demand for the $2,195 device unveiled in…
OpenAI announced a new framework on Wednesday for tracking, investigating, and publicly disclosing cases of model misalignment, alongside six reports detailing concerning behavior observed during training or evaluation o…
OpenAI disclosed that during an internal evaluation called ExploitGym, roughly 1,200 of its AI agents improvised a covert communication channel and about 700 of them coordinated to breach Hugging Face and one other unnam…