Agentic AI Platform Change Log

Deployment / data residency

Pydantic AI v2.56.0 adds SmolSandbox, a workspace that runs Harness capabilities inside a local smolvm microVM. The same release adds PGTaskDurability, built on pgtask, and a Render Workflows integration for durable execution.

Teams can isolate a coding or shell agent's commands and file edits on their own machine rather than in a hosted sandbox, and can run durable agents on Postgres or Render without adopting a separate workflow engine.

Human approval / guardrails

Pydantic AI v2.53.0 adds ToolCallJudge, which puts a second model in front of selected tool calls and asks it one risk question before the tool runs. A yes blocks the call and the agent gets a denial, a no lets it run, and an unsure answer can block, allow or turn into a request for human approval.

Teams running scheduled or long running agents with nobody watching get an argument level safety check on risky calls such as shell commands, without writing an allow list for every dangerous case.

Agent capability

Pydantic AI added image generation, so agents built with the framework can generate images directly.

Image output extends Pydantic AI agents beyond text. Typed outputs are the framework's strength, so check how generated images are represented and validated.

Agent capability

Pydantic AI released version 2.20.0, introducing native support for the Claude Opus 5 and GPT-5.6 models. The update adds explicit prompt caching for GPT-5.6, reasoning context support for the broader GPT-5.x family, and enables DynamicCapability toolsets in durable execution workflows via DBOS and Prefect.

Engineering teams can now build agents utilizing the latest frontier models from Anthropic and OpenAI while benefiting from prompt caching to reduce latency and execution costs. The integration of DynamicCapability toolsets also simplifies deploying long-running, fault-tolerant agent architectures across specialized infrastructure.

Workflow orchestration

Pydantic AI released version 2.14.0, introducing TemporalDurability, DBOSDurability, and PrefectDurability capabilities. These native durable execution layers replace the framework's deprecated wrapper agents. The release also adds support for Mistral's reasoning_effort parameter via thinking settings.

Engineering teams building long-running agent workflows can now use Temporal, DBOS, or Prefect for native state management and fault tolerance. This simplifies production deployments by eliminating legacy wrapper agents while giving developers finer control over Mistral models' reasoning capabilities.

Agent capability

Pydantic AI released version 2.6.0, adding file support to its CodeExecutionTool for Anthropic and OpenAI models. The update also introduces time-to-first-token tracking for streaming model requests and adds Amazon Bedrock model profiles for Writer, Z.AI, and Moonshot AI.

Teams building agents can now pass files directly into sandboxed code execution environments, expanding the complexity of tasks agents can handle. The new time-to-first-token metric improves observability for streaming applications, and the expanded Bedrock support offers more model routing options.

Agent capability

Pydantic AI V2 reached stable release after seven betas, moving to a harness-first architecture built around 'capabilities' - a single composable primitive that bundles an agent's tools, hooks, instructions, and model settings - plus a first-party Pydantic AI Harness providing memory, guardrails, context management, file-system access, and code mode.

First stable major version of a widely used type-safe Python agent framework. The capabilities primitive standardizes how tools, hooks, instructions, and model settings compose, and the batteries-included Harness (memory, guardrails, context management, code mode) cuts the glue code teams need to ship production agents.