Pydantic AI

First class Model Context Protocol support plus Agent2Agent and UI event stream standards, with built in capabilities for web search, web fetch, image generation, tool search, code execution and file access, and a composable capability system bundling tools, hooks, instructions and model settings into reusable units that can be built in house or installed as third party capability packages.

Sourcepydantic.dev/docs/ai/tools-toolsets/toolsetsread 2026-09-21

Agents compose into multi-agent systems through subagents and a dynamic workflow capability in which an orchestrator coordinates a catalog of sub-agents with fan-out, chaining and voting, and durable execution through Temporal, DBOS, Prefect, Restate and other backends lets long-running runs survive crashes and restarts.

SourcePydantic AI, pydantic.dev/docs/ai durable execution and harness pagesread 2026-09-21

A typed Embedder API generates embeddings across providers for semantic search and retrieval, and agents can pull context on demand through tools, but no index or vector store that Pydantic AI maintains over the customer's documents is documented, so retrieval is assembled per run by the developer; web search and web fetch reach the public web, not the customer's knowledge.

SourcePydantic AI, pydantic.dev/docs/ai/guides/embeddings and llms.txtread 2026-09-21

Deferred tools let a developer require approval before specific tool calls run, with the condition able to depend on the call's arguments, the conversation history or user preferences, and the Harness guardrails capability validates the prompt, tool calls and output with allow and block rules, so low and high risk actions can run under different autonomy modes.

SourcePydantic AI, pydantic.dev/docs/ai/tools-toolsets/deferred-tools and harness/guardrailsread 2026-09-21

The framework is a library with no hosted service or user layer; its documented controls are least-privilege tool access, through a FileSystem capability scoped to one directory tree with symlink-safe containment and a Shell capability with allow and deny controls and environment scrubbing. No supported identity providers or provisioning standards are documented, and no attestation covers the framework.

SourcePydantic AI, pydantic.dev/docs/ai/harness/filesystem and harness/shellread 2026-09-21

Built-in OpenTelemetry instrumentation records model calls and tool invocations as spans that flow to any OpenTelemetry backend the customer runs, so each step can be inspected. The trace store and views are the customer's chosen backend or the separately sold Pydantic Logfire, and the framework documents no audit log separate from runtime traces and no retention by plan.

SourcePydantic AI, pydantic.dev/docs/ai/capabilities/instrumentation and integrations/logfireread 2026-09-21

The Harness Memory capability gives each agent a persistent notebook of Markdown files it writes, reads, searches and deletes across runs, in a namespace resolved by application code so the model cannot reach another user's memory, with stores whose lifetime is stated, from process lifetime in memory to persistent file and SQLite stores; durable execution separately preserves run state across restarts.

SourcePydantic AI, pydantic.dev/docs/ai/harness/memoryread 2026-09-21

The agent runtime is an open source library that runs wherever the customer runs Python, with durable execution on backends the customer operates such as Temporal, DBOS and Prefect, so agent data stays on the customer's infrastructure.

SourcePydantic AI, pydantic.dev/docs/ai overview and durable execution pagesread 2026-09-21

Pydantic AI Harness ships complete agents alongside its capabilities: Coder, an autonomous coding harness with six tools and context management, and Researcher, a complete web-research harness that returns source-backed answers. Each does its own job without the other; the 30-plus capabilities themselves are parts the developer assembles.

SourcePydantic AI, pydantic.dev/docs/ai/harnessread 2026-09-21

Event driven workflows and UI event stream standards are supported with token by token streaming for interactive applications, but as a library the triggering surface and channel coverage are the host application's responsibility; no scheduler, webhook framework or channel matrix ships with the framework.

Sourcepydantic.dev/docs/ai/harness/gh-awread 2026-09-21

The developer chooses the model for each agent from a provider catalog that Pydantic AI describes as covering every model, with a documented path for custom models, and the Pydantic AI Gateway offers single-key access across providers or bring-your-own-key, so model choice sits with the customer.

SourcePydantic AI, pydantic.dev/docs/ai overview and overview/gatewayread 2026-09-21

Pydantic AI is an open source Python SDK developed in the open, with a documented public interface, a capability system for building and publishing custom capabilities, and agents that can be served to other systems over Agent2Agent and to editors over the Agent Client Protocol.

SourcePydantic AI, pydantic.dev/docs/ai overview, capabilities and harness/acp pagesread 2026-09-21

Pydantic Evals is a shipped, customer facing evaluation product rather than internal vendor tooling: create datasets, run evaluations, systematically test the performance and accuracy of the agentic systems you build, and track model performance over time from the CLI or visualized in Logfire, with the evaluation workflow integrated into the same OpenTelemetry trace as production runs.

Sourcepydantic.dev/docs/ai/evals/evalsread 2026-09-21

The Harness Playwright Browser capability gives an agent a real, stateful Chromium browser through async Playwright that it drives itself, navigating, clicking, typing, scrolling and extracting page text, with no separate account required; that is documented control of a real browser.

SourcePydantic AI, pydantic.dev/docs/ai/harness/playwrightread 2026-09-21