Agent InfrastructureAutomated release watch

Pydantic AI v2.62.0 (2026-10-26)

Pydantic AI v2.62.0 introduces mid-conversation thinking effort adjustment, new granular error categories, and SubAgent task control (backgrounding, steering, and cancellation). It also includes breaking changes for GraphBuilder replay order and MCP tool naming.

Human read

Why this signal matters

Pydantic AI released v2.62.0 with key feature updates and compatibility notes: - Mid-conversation thinking effort adjustment: Enabled via OpenAI config update and Anthropic per-message output config, tracked via ModelSettingsDeltaPart. - Error handling: Added ModelRequestRejectedError, ModelAuthenticationError, and ModelNotFoundError. - SubAgents task control: Lets models background, steer, answer, and cancel delegated SubAgent tasks. - MCP improvements: Hosted MCP capability tool names are now prefixed with the service name by default (configurable via tool prefix), and URL MCPToolsets without an ID default to their hostname. - Durable workflows: GraphBuilder broadcast branches now start in .to() order instead of reversed (potential replay breakage for paused workflows); run identity is preserved in Prefect capability replay; TemporalDurability allows running tools as Temporal child workflows. - Compaction and Memory: SummarizingCompaction stores summaries as CompactionPart instead of system prompts; added CompactOnRequest and Anthropic native compaction support.

Agent parse

Actionable summary

Pydantic AI v2.62.0 updates conversation management and durable execution. It adds mid-conversation thinking effort updates via ModelSettingsDeltaPart, new model error classes (ModelRequestRejectedError, ModelAuthenticationError, ModelNotFoundError), and SubAgent background task steering. Breaking/compatibility changes include starting GraphBuilder broadcast branches in .to() order, prefixing hosted MCP tool names…

Agent usefulness
77/100
Confidence
96%
Canonical data
JSON + Markdown
Next actions

What builders should check

  • Verify durable workflow replay if upgrading with in-flight workflows paused across GraphBuilder broadcast branches due to execution order changes.
  • Update MCP tool integration logic if relying on un-prefixed hosted MCP tool names, or set explicit tool prefixes.
  • Handle new model error types (ModelRequestRejectedError, ModelAuthenticationError, ModelNotFoundError) in exception handling blocks.
  • Migrate deprecated CodeMode monty sandbox url to CodeMode(monty sandbox=...) and Harness cache ttl settings.
Classification

Tags and routing

pydantic-aityped-agentspython
Related signals

Continue the thread