AI Agents Shift From Capability Race to Governance and Collaboration

Highlights

• Perplexity launches always-on Teammates, taking direct aim at OpenAI and Anthropic
• Microsoft's PDD Kit uses formal specifications to verify agent-generated code
• A Fortune 500 company found more than 10,000 unknown agents during a PoC

Work scenarios: always-on agents become the main battlefield

Perplexity released Teammates, an AI agent for work positioned as an always-on workspace member that can take on multiple roles at once and carry over a user's projects, skills and memory from existing Perplexity tools. It builds on the earlier Perplexity Computer, which can deploy agents to a user's devices to complete tasks, while Teammates emphasizes collaboration and "continuous work at scale." Users can assign different tasks to different Teammates to run in parallel, and reports compare it with OpenAI's Dots and Claude Duet. (Source: The Verge)

The significance of this thread lies in the shift in the competitive dimension: the selling point is no longer single-task capability, but memory, role division and long-term presence. For buyers, the evaluation focus will move from "can it do it" to "can it keep doing it without overstepping."

Interaction and execution: the entry point is changing, and so is the foundation

Meta announced that its Muse AI agent supports real-time voice conversation. The avatar at the top of the conversation moves in sync with what is being said, and users can directly ask the agent to adjust its tone or choose one in settings. In testing, voice responses were fast, but there were occasional pauses when querying information. (Source: The Verge)

More radical attempts are appearing on the hardware side: startup Aina is building a smart ring that argues users should not have to rely on a phone to control a virtual assistant, and can control an autonomous agent's actions with a single tap. (Source: Wired)

The execution layer is carried by infrastructure. NVIDIA notes that agents do more than generate answers: they search for information, query databases, write and test code, and take actions. Every step depends on the CPU, and tasks switch back and forth between GPU inference and CPU execution, so AI factories need CPUs that can handle these execution workloads. (Source: NVIDIA Blog) The same source also mentions that developers are combining frontier models with Omniverse libraries, using instructions to drive agents through building simulation applications, connecting physics and rendering, and validating scenes. (Source: NVIDIA Blog)

Trustworthiness: three gaps in verification, provenance and identity

Microsoft Research released PDD Kit to address the problem that agent-generated code has reached a scale where line-by-line human review is difficult: it turns developer intent into Verus formal specifications, proves that AI-generated code satisfies the specification for all inputs, supplements with tests where proof is not suitable, and lets developers review intent rather than every line. (Source: Microsoft Research Blog)

The gaps on the enterprise side are more concrete. MIT Technology Review described a typical chain: sales asks an agent about customer pricing, the agent pulls data from restricted pricing systems, internal communications and archived quotes, and the output looks accurate; a second agent uses it to draft a proposal, and neither handoff is checked by a human. (Source: MIT Technology Review) On identity, SailPoint said in remarks related to its Q2 FY2027 earnings that a Fortune 500 company discovered more than 10,000 previously unknown AI agents during a proof of concept; CB Insights used this to discuss security companies' next acquisition targets in the agent identity space. (Source: CB Insights Research)

Evaluation and commercialization: making "collaboration" measurable and billable

A new arXiv paper proposes SlackBench to evaluate agent performance in real research and engineering workflows: the environment reproduces real academic projects in a privacy-preserving way, includes multi-party conversations, evolving decisions, and code changes corresponding to GitHub repositories and Google Docs, with 140 queries in total, requiring agents to jointly reason between workplace communication records and repository states. (Source: arXiv cs.AI)

Commercialization support is also filling in. Clocknext appeared on Product Hunt, positioned as pricing and billing infrastructure for AI agents; Pushary for Teams focuses on showing who approved a team agent's actions; Maritime offers cloud computers that can be customized on demand for any agent to use. (Source: Product Hunt) These tools are still small, but they point to the same problem: when agents start doing work for people, authorization, billing and runtime environments need a separate layer.