What changed
Pydantic AI shipped v2.26.0 on August 7, 2026 with a notably more dynamic agent runtime. The headline change is support for hiding function tools until the model or app explicitly reveals them through tool search, load_capability, or ToolReturn.tools, as described in PR #7104. Instead of forcing every possible tool into the prompt upfront, an agent can now start with a smaller surface and expose more capabilities only when they become relevant.
The same release also adds first-party run cancellation through PR #6497 and promotes streaming run events into a public AgentRunEvents handle with cancellation and run-state access in PR #6498. Together, those changes move cancellation and tool-surface control out of application glue code and into the framework runtime itself.
Why it matters
This is a practical product shift for anyone building serious agents. Dynamic tool reveal helps reduce prompt bloat, lowers the chance that a model sees or calls the wrong tool too early, and makes approval-driven or discovery-driven workflows feel more natural. It also lines up with a broader agent design pattern: capabilities should be discoverable and staged, not dumped into context all at once.
First-party cancellation matters just as much. Long-running agent work is only trustworthy when developers can stop it cleanly, inspect state, and avoid orphaned background activity. By exposing cancellation and event access as official runtime features, Pydantic AI is making control surfaces that production teams actually need part of the default developer experience.
What builders should watch
Builders should watch whether Pydantic AI now becomes a stronger fit for agents that need gated tools, progressive capability loading, or user-controlled interruption. The deeper signal is architectural: agent frameworks are starting to treat capability exposure and cancellation as core product behavior, not optional wrapper logic. That makes the runtime itself a larger part of the product decision.