Framework

Pydantic AI errors and fixes: 12 guides

Real Pydantic AI failures, each with the error you'll see, the cause and steps that worked. 12 of 12 fixed upstream. Search within Pydantic AI.

Fixed: 12

Pydantic AIFixed (verified)
AG-UI tool calls point to a parent message that was never announced
For a response with only tool calls, the AG-UI event stream emits TOOL_CALL_START with a parentMessageId that no TEXT_MESSAGE_START ever announced.
Last checked Oct 17, 2026#messaging#state1 steps →
Pydantic AIFixed (verified)
Concurrent runs sharing an MCP server fail with a cancel scope RuntimeError
Running several agent runs or evals at once against the same MCP server toolset eventually fails with RuntimeError: Attempted to exit cancel scope in a different task than it was entered in.
Last checked Oct 17, 2026#race#transport2 steps →
Pydantic AIFixed (verified)
Custom httpx client given to an MCP server is closed after one run
After the first agent.run(), the httpx.AsyncClient passed to MCPServerStreamableHTTP is closed, and the next run fails with RuntimeError: Cannot reopen a client instance, once it has been closed.
Last checked Oct 17, 2026#transport1 steps →
Pydantic AIFixed (verified)
new_messages() loses messages when a history processor trims the current run
With a history processor that keeps only the last few messages, result.new_messages() returns just the last few messages of the run instead of everything the run added.
Last checked Oct 17, 2026#state1 steps →
Pydantic AIFixed (verified)
new_messages() returns the user prompt passed in message_history
Calling agent.run() with message_history and no user_prompt makes result.new_messages() include the request from that history, even though it was input, not something the run produced.
Last checked Oct 17, 2026#state#regression2 steps →
Pydantic AIFixed (verified)
prepare_tools callback returning None silently hides every tool
After a prepare_tools callback returns None, the agent runs with zero tools and answers in prose, with no exception, log or trace entry.
Last checked Oct 17, 2026#tool-calls#error-messages2 steps →
Pydantic AIFixed (verified)
run_sync() freezes when a sync output function calls another agent
An agent run with run_sync sometimes hangs forever when its output function calls a second agent with run_sync.
Last checked Oct 17, 2026#race#tool-calls2 steps →
Pydantic AIFixed (verified)
Sliding window compaction drops the current request when keeping one message
With SlidingWindowCompaction(keep_messages=1, receipts=True), every time it trims, the model gets only the compaction receipt and never sees the prompt it is supposed to answer.
Last checked Oct 17, 2026#state#config2 steps →
Pydantic AIFixed (verified)
Streamed run stops right after a tool call when the model sends empty text first
With some local models, an agent run that uses tools stops right after the tool is called and returns nothing.
Last checked Oct 17, 2026#parsing#tool-calls1 steps →
Pydantic AIFixed (verified)
system_prompt is skipped when the agent runs through a UI adapter
An agent served through AGUIAdapter, VercelAIAdapter or to_web runs without its system_prompt, and functions decorated with @agent.system_prompt never run.
Last checked Oct 17, 2026#config#state2 steps →
Pydantic AIFixed (verified)
Temporal retries an MCP tool forever when it raises ModelRetry
In a Temporal agent, an MCP tool that raises ModelRetry, such as a fetch that hits a 404, fails the activity, and Temporal retries it endlessly instead of handing control back to the agent.
Last checked Oct 17, 2026#loop#tool-calls1 steps →
Pydantic AIFixed (verified)
web_fetch_tool ends the run on a URL with an invalid hostname
When the model asks web_fetch_tool for a URL whose hostname fails IDNA parsing, the run ends with httpx2.InvalidURL instead of letting the model try again.
Last checked Oct 17, 2026#tool-calls#error-messages1 steps →

Get 3 fixes every Monday in the Weekly Fix →

Pydantic AI errors and fixes: 12 guides

Fixed: 12

AG-UI tool calls point to a parent message that was never announced

Concurrent runs sharing an MCP server fail with a cancel scope RuntimeError

Custom httpx client given to an MCP server is closed after one run

new_messages() loses messages when a history processor trims the current run

new_messages() returns the user prompt passed in message_history

prepare_tools callback returning None silently hides every tool

run_sync() freezes when a sync output function calls another agent

Sliding window compaction drops the current request when keeping one message

Streamed run stops right after a tool call when the model sends empty text first

system_prompt is skipped when the agent runs through a UI adapter

Temporal retries an MCP tool forever when it raises ModelRetry

web_fetch_tool ends the run on a URL with an invalid hostname