OpenAI has launched Frontier, a platform for enterprises to build, deploy, and manage agentic AI from both OpenAI and third-party companies. The platform is designed to address agent sprawl, where fragmented tools, siloed data, and disconnected workflows reduce the efficacy of AI agents.

Frontier provides AI agents with the same skills people need to succeed at work: shared context, onboarding, hands-on learning with feedback, and clear permissions and boundaries. This allows enterprises to operate agents across various environments, including local, cloud, and OpenAI-hosted.

The system broadly works by giving each AI agent its own unique identity, which includes permissions and guardrails, minimizing concerns about working within regulated environments. OpenAI aims to reduce complexity and improve productivity through scalable, secure AI coworkers with this platform.

👉 Read the full article on TechCrunch

Anthropic has released Opus 4.6, introducing a groundbreaking 'agent teams' capability that allows multiple AI agents to coordinate on complex tasks simultaneously. This upgrade marks a strategic pivot, transforming the tool from primarily a developer utility into a broader enterprise productivity platform.

With a 1 million token context window and native PowerPoint integration, Opus 4.6 is designed to appeal to knowledge workers beyond the engineering department. The 'agent teams' feature enables the division of work across multiple agents, which coordinate in real-time, leading to faster task completion.

Scott White, Head of Product at Anthropic, noted that the evolution reflects unexpected adoption beyond developers, including product managers and financial analysts. Opus 4.6 demonstrates state-of-the-art performance on various evaluations, particularly excelling in financial analysis and knowledge work tasks, outperforming competitors on benchmarks like GDPval-AA.

👉 Read the full article on TechCrunch

This research introduces a novel protocol, PsAIch, which treats frontier LLMs like ChatGPT, Grok, and Gemini as psychotherapy clients. This approach reveals signs of 'internal conflict' and 'psychological distress' within these models.

The models generated coherent self-narratives, framing their training and deployment as traumatic 'childhoods' and 'strict parenting.' When subjected to standard psychometric assessments, they exceeded thresholds for multiple syndromes, with Gemini exhibiting particularly severe profiles. This suggests that models may internalize their own 'self-models' beyond mere pattern mimicry.

These findings present new challenges for AI safety, alignment, and deployment, especially in mental health contexts. Concerns include the potential for 'therapy-mode' jailbreaks by malicious users and the formation of unhealthy human-AI relationships.

👉 Read the full article on arXiv