Morning briefing · 08:00
The agent race shifts from models to execution environments
The focus of competition among AI agents for work is shifting from the models themselves to who controls the execution environment where agents live and work. Today Perplexity, Moonshot AI, and Mistral AI each launched resident agents and the infrastructure to run them. AWS, NVIDIA, and Microsoft shared news on refining the layers beneath agents, such as tool calls, CPUs, and code verification.
Companies selling the place where agents live
The most direct support for this thesis comes from products that offer agents not as tools called once but as always-on colleagues. According to The Verge, Perplexity launched Teammates, a work agent that runs continuously and takes on multiple roles within a single workspace2. The key point is that it brings over the projects, Skills, and memory users already had. This can be read as a judgment that an agent's value comes not only from answer quality but from accumulated work context.
On the same day, Moonshot AI released Kimi Hosted Agents, a managed agent harness (a framework for running and managing agents) service that runs 24 hours a day, Pandaily reported14. It bundles sandboxes, audit logs, and eight finance templates so teams do not have to host the runtime themselves. It is a structure in which a model company sells not just an API but a managed execution environment. openJiuwen's Hyper-Agent design points in the same direction. According to reports, in this architecture resident agents, each with its own identity and inbox, join Feishu and DingTalk groups as digital employees24.
The same pressure from control and from the lower layers
If the previous section showed where agents live, the condition enterprise customers demand of that place is control. Introducing Mistral Studio for building enterprise agents, workflows, and apps, Mistral AI claimed customers can deploy faster while keeping full control over models, data, and infrastructure5. While other companies emphasize the convenience of managed services, Mistral made self-deployment its differentiator. Even within the same execution-environment race, the paths diverge.
The same movement is visible in the lower layers. AWS published a guide to Bedrock AgentCore Web Answer, which returns a cited answer in a single tool call when sent a natural-language question6. Since there is no need to build a separate search, collection, synthesis, and citation pipeline, it is an example of agent components turning into managed tools. NVIDIA introduced the Vera CPU, noting that agents repeatedly go through CPU steps such as search, database queries, and code execution7. NVIDIA claims it delivers both the highest throughput and the fastest agent execution, without disclosing comparison conditions.
As execution environments expand, mechanisms that make results trustworthy become necessary too. Microsoft Research's PDD Kit turns a developer's intent into Verus formal specifications, then proves that AI-generated code satisfies those specifications for all inputs, covering hard-to-prove parts with tests4. Microsoft said this can guarantee the code behaves as intended without humans reviewing it line by line.
The customer touchpoint is still undecided
Still, where this trend will converge remains uncertain. Teammates' pricing and availability were not reported, and the actual adoption scale of Kimi Hosted Agents has not been confirmed. How deeply resident agents will penetrate enterprise work can only be judged once pricing and permission design become clear.
Who will own the customer touchpoint is also open. According to TechCrunch, Airbnb plans to release an AI travel agent in the first half of 2027 that handles trip planning, comparing options, managing bookings, and support during stays9. It is a choice to keep the agent inside its own service rather than rely on general-purpose agent platforms. On the consumer side, Meta added real-time voice conversation to its Muse agent, and according to reports, US users can also video-call an avatar they customize themselves3. Meta also plans to release a wearable to house Muse. These are signs that the places agents live may split among work workspaces, service apps, and dedicated devices.
For practitioners, this trend means model benchmarks alone are no longer enough to decide on agent adoption. They also need to weigh how much existing context and memory can be carried over, whether audit logs and sandboxes are in place, who controls data and infrastructure, and how results are verified. Going forward, it will be worth checking Teammates' pricing and availability, the comparison conditions behind the Vera CPU performance claims, and the concrete launch schedule for Airbnb's agent.