Pydantic AI vs LangGraph: typed agents or an explicit graph
Pydantic AI is a type-safe Python agent framework built around capabilities, with durable execution through engines such as Temporal and DBOS and OpenTelemetry tracing, while LangGraph is an explicit state-graph runtime in Python and TypeScript, so Pydantic AI fits code-first agents with structured outputs and LangGraph fits workflows whose control flow needs to be drawn as a graph.
Checked
Pydantic AI v2 shipped on 23 June 2026 and changed the framework's core: capabilities became the main primitive, and memory, guardrails and code mode moved into a separate Pydantic AI Harness. Comparisons written before v2 describe a different framework.
LangGraph takes the opposite starting point. The agent is a graph whose nodes and edges are declared explicitly, with persistence and human approval built into the runtime. The choice is mostly about where a team wants control flow to live: in ordinary typed Python, or in a graph.
What changed in Pydantic AI v2
Capabilities
Capabilities became the core primitive of an agent.
Harness split
Memory, guardrails and code mode moved to a separate Pydantic AI Harness.
Providers
Non-frontier providers are now opt-in, and openai: models use the Responses API.
Release cadence
The no-breaking-changes window between majors shrank from six months to three, after v1 in September 2025.
Durability and observability
Pydantic AI
Durable execution through Temporal, DBOS, Prefect, Restate, AWS Lambda, Absurd, pgtask and Apache Airflow. Built-in OpenTelemetry instrumentation following the GenAI semantic conventions, sent to Pydantic Logfire or any OpenTelemetry-compatible backend.
LangGraph
Durable execution and persistence in the runtime itself, human-in-the-loop editing of state, and debugging through LangSmith.
Which fits which job
Pydantic AI fits
Python teams that already use Pydantic and FastAPI, agents whose outputs must validate against a schema, and teams that want tracing on an open standard.
LangGraph fits
Multi-step workflows with branches, retries and approval points, and teams that need the same runtime in TypeScript.
How Motiontic approaches it
Motiontic's engineers are framework-neutral. They choose per project after a short spike on the client's real task, keep prompts, tool definitions and evals outside the framework so a later switch is a refactor rather than a rewrite, and hand the eval suite to the client's team so the next framework decision can be measured.
Pydantic AI and LangGraph, figures retrieved 30 October 2026
| Pydantic AI | LangGraph | |
|---|---|---|
| Core model | Typed agents built from capabilities | Explicit graph of nodes over shared state |
| Licence | MIT | MIT |
| Languages | Python | Python, JavaScript / TypeScript |
| Latest version | 2.63.0 (28 Oct 2026); 2.0 on 23 Jun 2026 | 1.2.18 (26 Oct 2026); 1.0 on 17 Oct 2025 |
| Durable execution | Via Temporal, DBOS, Prefect, Restate and others | Built into the runtime |
| Observability | OpenTelemetry; Logfire or any OTel backend | LangSmith |
| A2A support | Listed as an adopter | Listed as an adopter |
| GitHub stars | 20,959 | 44,172 (Python repo) |
| PyPI downloads, last week | 1.71M (pydantic-ai) | 13.76M |
Common questions
What changed in Pydantic AI v2?
Released on 23 June 2026, v2 made capabilities the core primitive, moved memory, guardrails and code mode into a separate Pydantic AI Harness, made non-frontier providers opt-in and switched openai: models to the Responses API.
Does Pydantic AI support durable execution?
Yes, through external engines. Its docs list Temporal, DBOS, Prefect, Restate, AWS Lambda, Absurd, pgtask and Apache Airflow.
Does Pydantic AI require Logfire?
No. Logfire support is optional. Pydantic AI follows the OpenTelemetry semantic conventions for generative AI and can send traces to any OpenTelemetry-compatible backend.
Is Pydantic AI simpler than LangGraph?
For a single agent with typed outputs, usually yes, because control flow stays in ordinary Python. For a workflow with many branches and approval steps, LangGraph's explicit graph can be easier to reason about.
Related
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
LangChain vs LangGraph
The agent layer and the runtime underneath it, compared after the 1.0 releases.
Claude Agent SDK vs LangGraph
Claude Code's harness as a library compared with a model-neutral graph runtime.
Agent observability
Step-by-step traces of what an agent did, why, and what it cost.
OpenTelemetry GenAI semantic conventions
Standard gen_ai.* names for agent and model telemetry, still at Development status.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.