Morning briefing · 08:00
Agents move into apps and professional tools
Where AI agents do their work is shifting from a separate chat window into the apps and professional tools people already use. Today several companies presented agents working inside professional tools for drug discovery, hardware design, and game engines. An operating system's voice assistant has also started operating external apps directly. At the same time, there are moves to manage the connectors that link agents to apps as their own category.
Agents inside professional tools
Today's clearest signal is that agents have started working inside specialized software for specific fields. Google DeepMind added cheminformatics-specific Skills (bundles of instructions and tools needed for particular tasks) to Google Antigravity. It also showed how to use them to handle drug discovery tasks with agents10. The examples went beyond compound search. They also included proposing new macrocycle candidates with more rigid structures. OpenAI published a post describing how engineers use ChatGPT and Codex to speed up hardware development work. The tasks covered include CAD, PCB design, and simulation9.
An announcement in the same direction came from China. Kimi released a game development solution that bundles the K3.1 model with the Kimi Code and Kimi Work agents. The agents directly edit Unity, Unreal Engine, and Blender projects. A case in which they ran continuously for 20 hours without human intervention was also presented6. The three cases span different fields but share a common thread. Rather than building new tools, they attach agents to the programs and work files practitioners already use. Domain knowledge is layered on through Skills or dedicated configurations. The approach refined in coding agents can be seen spreading to chemistry, mechanical and electronic design, and game production.
The trend spreads to everyday apps and enterprise connectors
The trend seen in professional tools is also visible in consumers' smartphones and enterprise systems. According to The Verge, Apple's iOS 27.3 update, which began rolling out on Tuesday, lets developers integrate Siri AI's new capabilities into any app1. Siri understands the user's personal context and on-screen content. It then carries out actions inside third-party apps in response to natural-language requests. App developers can no longer design only for users opening their app directly. They also need to design for users arriving through the operating system's assistant. The DeepSeek Harness team's 'Creator Mode' does the same thing from another direction. According to the report, when a user describes the feature they want, the AI writes the plugin code and also handles installation and checking that it works13.
The more apps and systems agents connect to, the more those connections become something to manage. GitHub Dependabot has started supporting Agent Package Manager (APM). When a new version of an agent Skills, prompt, or instructions package used by a repository is released, it automatically opens an update PR11. In effect, agent configuration files have entered the framework of version and supply chain management, like code dependencies. AWS published how to set up on-behalf-of (OBO, delegating permissions to act for a user) token exchange with Microsoft Entra ID in Amazon Bedrock AgentCore Gateway22. In this approach, the agent calls the ServiceNow API while preserving the identity of the calling user rather than a service account. Version management and per-user permissions are requirements enterprises check first before agents touch real business systems.
The gap between working longer and being trustworthy
But agents working in more tools for longer does not mean their results can be trusted. A team led by Professor Zhou Hongyu of Tsinghua University analyzed the diagnostic performance of large reasoning models using real patients' electronic medical records. The team reported that longer reasoning improved accuracy but also increased the risk of adding incorrect diagnoses or giving different diagnoses for the same patient12. The study covers only the medical field. Still, it provides grounds for separately examining accuracy and consistency when looking at cases that emphasize task length, such as long unattended runs.
Tools to evaluate and verify this are only now being strengthened. Ai2 introduced 'StateBench', which tests how well language models follow state changes in long documents. The same day, 'CWE-bench v2', based on security weakness types (CWE), was released with improved scope, quality, and transparency18. The long-context tracking and secure code evaluation covered by the two benchmarks are considered weak spots in practical agent work. TechCrunch reported that Anthropic is building a system called 'FlashProve'2. It would let anyone check whether an AI response actually came from the model in question. The technical approach and release schedule have not yet been disclosed. Still, the more agents move across apps and tools, the greater the need for mechanisms to confirm which model did what.
The takeaway from today's news for practitioners is that model performance alone is no longer enough to decide on agent adoption. You need to decide which tools to attach agents to. You also need to weigh how to version Skills and instructions, whose permissions the agent will act under, and how to measure the consistency of its results. There are three things to watch going forward. First, how far the number of apps actually participating in Siri integration grows. It is also not yet known what verification method FlashProve uses or whether it applies to other companies' models. Finally, it remains to be seen whether long-running cases will be published along with accuracy and consistency metrics.