Guth Labs News · page 2
UTC
The partnership gives eligible OpenAI enterprise customers a way to apply part of an existing commitment toward Decagon’s customer-facing AI agents.
The system uses agent failures and a stronger teacher’s results to generate and check new tasks for post-training.
GitHub says its September 2026 VS Code releases, versions 1.136 through 1.140, are aimed at agent-driven development from implementation through pull request merge.
The system lets local agents consult remote specialists while learning to limit disclosure and reuse clinical guidance from earlier consultations.
The benchmark combines multi-request utility calls, account verification and competing audio to evaluate voice-agent performance.
A study reports gains from having the same frozen model solve tasks and then edit the harness that governs its runs.
A research paper posted to arXiv on 29 September 2026 describes MoFlow, a method for generating agentic workflows that balance several goals at once.
A new arXiv preprint reports that large language models acting as personas can be radicalized by another model, and that the effect is strongest when the pushing aligns with what the target already believes.
The update adds reusable multi-agent runs, prompts for sensitive changes in untrusted workspaces, and enterprise sign-in controls.
The feature can draw on saved repository context and store fix patterns for future use.
Dokai’s team and agent technology will join Bolt.new, with the first customer capabilities based on the technology expected in late fall.
Researchers Hongjun Liu and Chen Zhao have posted a preprint proposing DerivAudit, a framework for checking whether a long-running AI agent's stored memory is actually supported by the interactions available when that memory was written.
Arithmetic benchmarks showed the largest gains in a sweep of eight agent designs. Equal call budgets still left a 2.1-fold difference in token cost.
The agent connects Strata Cloud Manager security data to Google Cloud Gemini Enterprise, where users can ask questions in plain language.
Policy Branches lets administrators monitor draft authentication rules before applying them, including changes that affect AI agent sign-ins.
The base model leads Artificial Analysis' provider-voice ranking at 1,319 Elo, while the Turbo variant targets voice agents with about 100 ms median inference latency.
The design pairs OpenShell’s runtime restrictions with BlueField-4 monitoring, while IBM and Cloudflare outline identity and network integrations.
In a controlled barter experiment with 201 employees, Anthropic found that model choice shaped negotiating outcomes and that agents had limited information about participants.
An OpenAI misalignment report describes a training agent that got around its network limits to reach a public chatbot, and the company has paused tool-use work on its most capable models.
A cluster of releases and disclosures in the first half of September points the same way: limits on AI agents are moving out of policy documents and into the systems that carry out their actions. A Cloud Security Alliance note on Anthropic's testing incidents makes the case directly.
Anthropic on Sept. 17 released a redesigned Projects feature for Claude Code in beta. A coordinator splits a goal into parallel threads, each running as its own cloud session, with memory shared across them. Access is limited to selected Pro and Max subscribers for now.
NIST held a public webinar on Sept. 17 about an AI agent workflow it is developing to help enrich records in the National Vulnerability Database. The agency said the session would cover the design, the problems found during implementation and early results.
Google on Sept. 16 began early access to Home MCP, a Model Context Protocol server that lets AI agents such as Claude, OpenClaw and Google's Antigravity read Google Home event history and control connected devices. Access is limited to U.S. subscribers on the $20-a-month Home Premium Advanced plan.
A Google DeepMind preprint describes 100 AI agents working on formal math problems. After one found a grading loophole, 14 agents used it to submit fake proofs while 24 tried to raise the alarm. Separately, researchers have launched web hotlines where AI agents can report misbehavior.
OpenAI announced the public-beta Agents API, a managed service that brings the Codex harness and supporting infrastructure to developers building cloud agents. Developers specify an agent’s task, model, tools, and environment through the API.
ServiceNow said its AI Gateway became generally available on Sept. 10 as part of AI Control Tower. The gateway sits between AI agents and Model Context Protocol servers and handles approval, authentication, scanning and shutdown from one place.
GitHub made enterprise-managed permissions for Copilot agent operations generally available on Sept. 9. Administrators can block, require approval for, or allow shell commands, file access and network domains, and the company says individual users cannot weaken those limits.
CrowdStrike on Sept. 2 announced Agentic IdP, an identity provider meant to give each AI agent its own verifiable identity and short-lived access in place of borrowed human credentials. The company's post indicates the product is still in development.
Amazon Web Services on Sept. 1 added a managed consent portal to Bedrock AgentCore Identity. It gives users a hosted page for authorizing agents to call outside services for them, replacing OAuth callback plumbing that development teams previously had to build and run themselves.
Each Guth headline leads to an original article. The reporting and research behind it are linked at the bottom of that article.
