NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #1934 most downloaded on PyPI
The official capability library and harness for Pydantic AI
Latest stable
0.66.0
Last release 18 days ago
04 Nov 2026
Ships on a steady schedule
a new release about every 9 days
Most releases are documented
notes for 41 of 56 stable releases
Nothing withdrawn
no release was ever pulled
7 months old
56 releases · first in 2026
Drag across the ruler to compare, or tap a release to open it.
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Your coding agent can read these notes before it upgrades. Set up the MCP server →
One column per month.
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Add hosted MCP tools with Grain by @adtyavrdhn in #1088
Grain by @adtyavrdhn in #1088DayAI by @adtyavrdhn in #1089PostHog by @adtyavrdhn in #1090Pylon by @adtyavrdhn in #1091pydantic-monty>=1 and add web sockets support by @adtyavrdhn in #524Model.context_window, and ask for a summarizer model on a text-less run by @DouweM in #1101Coder delegate to the agent it is bound to, via SubAgents(include_self=True) by @DouweM in #1061Full Changelog: v0.35.0...v0.36.0
Add hosted MCP tools with LogfireMCP by @adtyavrdhn in #806
LogfireMCP by @adtyavrdhn in #806GitHub by @adtyavrdhn in #799Linear by @adtyavrdhn in #793Notion by @adtyavrdhn in #797GoogleWorkspace by @adtyavrdhn in #791Slack by @adtyavrdhn in #808Ordinal by @adtyavrdhn in #861CodeMode results as repr instead of aborting the run by @mpfaffenberger in #1077Full Changelog: v0.34.0...v0.35.0
Clarify Code Mode speculation setup, safety, and metrics by @mpfaffenberger in #1022
FileSystemToolset without replacing local write safeguards by @mpfaffenberger in #1031Shell.max_file_bytes by @mpfaffenberger in #1030Advisor accept output_type for structured consultations by @mpfaffenberger in #1028Shell.max_file_bytes in configuration summaries by @mpfaffenberger in #1041CodeMode eagerly scans multiple statements by @mpfaffenberger in #1044Coder output limits a stable identity for durable execution by @mpfaffenberger in #1043Full Changelog: v0.33.0...v0.34.0
Add BackgroundTools for concurrent tool execution by @DouweM in #222
BackgroundTools for concurrent tool execution by @DouweM in #222PAI_AGENT spec omit the model and raise the pydantic-ai floor to 2.44.0 by @DouweM in #953speculate) for CodeMode by @mpfaffenberger in #699Full Changelog: v0.32.0...v0.33.0
Warn on the deprecated single-key get / add of InMemorySpendStore and RedisSpendStore by @xixi-box in #747
file_system events by @mpfaffenberger in #853Coder around six tools and persistent shell commands by @mpfaffenberger in #870Coder on Terminal-Bench with Harbor by @mpfaffenberger in #872Coder from FileSystem and Shell by upstreaming its tools into them by @mpfaffenberger in #895gh-aw engine's agent when a workflow configures observability.otlp by @DouweM in #862gh-aw engine experimental by @DouweM in #891FileSystem a max_read_chars budget so Coder reads page without skipping lines by @Quiyyy in #892FallbackCompaction by @mpfaffenberger in #913RepairToolArguments as a standalone capability by @mpfaffenberger in #917Coder instructions to engineering guidance by @mpfaffenberger in #916FileSystem discovery paths with cwd by @Quiyyy in #943get/add of InMemorySpendStore and RedisSpendStore by @xixi-box in #747Coder's plumbing last and fix the Coder docs drift by @DouweM in #949WarnOnCacheBusts, across the runs that continue it by @DouweM in #825Full Changelog: v0.31.0...v0.32.0
Drop the unfulfilled 0.28.0 removal date from SpendStore deprecations by @mpfaffenberger in #841
eager streamed execution to CodeMode by @mpfaffenberger in #746PlanStore get_items and set_items docstrings by @mpfaffenberger in #843tool_return_max_chars to SummarizingCompaction by @mpfaffenberger in #842SpendStore deprecations by @mpfaffenberger in #841step_persistence docs by @mpfaffenberger in #840agent_docs, ported from core by @mpfaffenberger in #835Truncate by @Quiyyy in #836Planning cache breakpoint on the last durable user content by @pydanty in #833FileSystem content hashes agree across read_file, write_file, and edit_file by @mpfaffenberger in #847SnapshotHistorySource.run_history duplicating messages with serializing stores by @mpfaffenberger in #848pydantic-ai 2.40.0 by @mpfaffenberger in #851DelegationStartEvent and DelegationEndEvent from SubAgents by @mpfaffenberger in #855aws-lambda docs install line, which named a bedrock extra that does not exist by @mpfaffenberger in #854pip and uv installation commands throughout the Harness docs by @mpfaffenberger in #858CodeMode sandbox the value shape its own stubs declare by @bohdanhr in #681Full Changelog: v0.30.0...v0.31.0
Deprecate on_spend , on_usage , on_fire , and PlanEventEmitter toward capability event subscriptions by @DouweM in #714
MCPToolsetClient from pydantic_ai_harness.stackone._toolset so CI lint passes again by @mpfaffenberger in #725media docs as Step Persistence plumbing, not a capability by @mpfaffenberger in #751Agent in SummarizingCompaction, ToolOutputLimits, SystemReminders, Advisor and PydanticAIChatModel so Logfire attributes their runs by @mpfaffenberger in #774gather teardown in guardrails parallel mode and MontyExecutor by @mpfaffenberger in #758TrajectoryJudge: review the live run with a second model on a cadence and steer it mid-run by @mpfaffenberger in #724on_spend, on_usage, on_fire, and PlanEventEmitter toward capability event subscriptions by @DouweM in #714FileSystem traversal events and subscribe RepoContext with @on_event, deprecating traversal_tool_names sniffing by @DouweM in #712youdotcom>=3.1.2 and pass attribution kwargs to You() so outbound requests carry an X-Client-Info header by @tyler5673 in #786pydantic-ai gh-aw engine definition is now hosted here, and PAI_AGENT points a workflow at an agent your repository defines by @dsfaccini in #708Truncate.max_chars by @Quiyyy in #816Full Changelog: v0.29.1...v0.30.0
Handle instruction updates in token estimates and conversation search by @dsfaccini in #815
pydantic-monty floor to 0.0.23 by @adtyavrdhn in #813Full Changelog: v0.29.0...v0.29.1
Floor on released pydantic-ai 2.38.0 by @DouweM in #780
pydantic-ai 2.38.0 by @DouweM in #780SummarizingCompaction stream nested summary requests by @JayGhiya in #620media marker keys and match the escape stash by shape by @dsfaccini in #700SubAgents and StackOne the ids that decide it by @DouweM in #767Full Changelog: v0.28.1...v0.29.0
Make StepPersistence replay-safe under durable execution by @DouweM in #749
StepPersistence replay-safe under durable execution by @DouweM in #749SpendLimits replay-safe under durable execution by @DouweM in #748pydantic-ai 2.37.0 and drop the version straddle by @DouweM in #769RunCancelled past contain_errors by @mpfaffenberger in #754tests/logfire_variables so a local .env cannot reach the real Logfire API by @mpfaffenberger in #752Advisor's default id on the class by @DouweM in #776Full Changelog: v0.28.0...v0.28.1
Separate instruction parts from different capabilities with a blank line by @DouweM in #698
pydantic-ai 2.36.0 and drop the version straddle by @DouweM in #743Further reading section to the PlaywrightBrowser docs and README by @dsfaccini in #654SummarizingCompaction summarization a durable operation by @DouweM in #707Full Changelog: v0.27.0...v0.28.0
Allow compaction summaries to override model settings by @thomwebb in #650
spend doc snippets for undefined names, so they stay runnable on the 3.10 floor by @Neallin-917 in #701Full Changelog: v0.26.0...v0.27.0
Apply one response to every budget window as one BatchSpendStore operation, deprecate SpendStore for removal in 0.28.0, and hash-tag the RedisSpendSto…
Shell spawn failures to the model instead of aborting the run by @dsfaccini in #623BatchSpendStore operation, deprecate SpendStore for removal in 0.28.0, and hash-tag the RedisSpendStore keys by @dsfaccini in #548SpendLimits, and correct the durability, composition and UsageLimits claims in the shipped docs by @dsfaccini in #547Full Changelog: v0.25.0...v0.26.0
Allow release gates to pin harness-ref by @dsfaccini in #652
harness-ref by @dsfaccini in #652pydantic-ai against dependencies it does not support by @dsfaccini in #659pydantic-ai-slim>=2.33.0 so anthropic resolves to the HTTPX2 generation by @dsfaccini in #664UsageLimits to nested summaries by @dsfaccini in #665instructions on SummarizingCompaction by @mpfaffenberger in #669Full Changelog: v0.24.0...v0.25.0
Add a serializer seam and indented_json / json_lines presets so ToolOutputLimits spills page by line by @adtyavrdhn in #638
serializer seam and indented_json/json_lines presets so ToolOutputLimits spills page by line by @adtyavrdhn in #638FileSystem tools from leaking absolute host paths to the model by @ada-zl5825 in #628YouSearch and YouResearch capabilities backed by the You.com APIs by @itsakhilyou in #646FileSystem failures to the model instead of aborting the run by @dsfaccini in #612PlaywrightBrowser capability driving Chromium through async Playwright by @dsfaccini in #420write_file from blocking on FIFOs by @samrusani in #613Full Changelog: v0.23.0...v0.24.0
Keep the ManagedPrompt baggage snapshot valid when the resolution span nests under the run span by @adtyavrdhn in #641
ManagedPrompt baggage snapshot valid when the resolution span nests under the run span by @adtyavrdhn in #641FallbackCompaction for failure-based strategy chains by @mpfaffenberger in #644Full Changelog: v0.22.0...v0.23.0
Fix protected pattern semantics for read-only walkers by @dsfaccini in #364
PromptInjectionDefender for indirect prompt injection by @shashi-stackone in #439CodeMode from folding shell command tools into run_code by @kaxil in #593compaction doc examples to claude-sonnet-5, claude-opus-5, gpt-5.6-luna/terra, and gemini-3.6-flash by @mpfaffenberger in #639Full Changelog: v0.21.0...v0.22.0
Front pages + first harnesses: Coder and Researcher, one capability index across core and Harness by @DouweM in https://github.com/pydantic/pydantic-a
Full Changelog: https://github.com/pydantic/pydantic-ai-harness/compare/v0.20.0...v0.21.0
Full Changelog: v0.20.0...v0.21.0
Fix Planning cache-point order by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/570
Planning cache-point order by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/570SpeechPart and AbstractModel run models from pydantic-ai 2.28 by @DouweM in https://github.com/pydantic/pydantic-ai-harness/pull/586hatchling below 1.32 for release builds by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/595Full Changelog: https://github.com/pydantic/pydantic-ai-harness/compare/v0.18.1...v0.20.0
Planning cache-point order by @dsfaccini in #570SpeechPart and AbstractModel run models from pydantic-ai 2.28 by @DouweM in #586hatchling below 1.32 for release builds by @dsfaccini in #595Full Changelog: v0.18.1...v0.20.0
Read deferred-tool hiddenness from visibility, not absence by @DouweM in https://github.com/pydantic/pydantic-ai-harness/pull/573
ToolAvailabilityDeltaPart in the compaction estimator and conversation-search index by @adtyavrdhn in https://github.com/pydantic/pydantic-ai-harness/pull/579Full Changelog: https://github.com/pydantic/pydantic-ai-harness/compare/v0.18.0...v0.18.1
ToolAvailabilityDeltaPart in the compaction estimator and conversation-search index by @adtyavrdhn in #579Full Changelog: v0.18.0...v0.18.1
feat(browseruse): delegate open-ended web tasks to a browser-use agent by @DEENUU1 in https://github.com/pydantic/pydantic-ai-harness/pull/419
Full Changelog: https://github.com/pydantic/pydantic-ai-harness/compare/v0.17.0...v0.18.0
feat(spend): add SpendLimits for spend and token budgets across windows and processes by @DEENUU1 in https://github.com/pydantic/pydantic-ai-harness/p
SpendLimits for spend and token budgets across windows and processes by @DEENUU1 in https://github.com/pydantic/pydantic-ai-harness/pull/474Full Changelog: https://github.com/pydantic/pydantic-ai-harness/compare/v0.16.0...v0.17.0
SpendLimits for spend and token budgets across windows and processes by @DEENUU1 in #474Full Changelog: v0.16.0...v0.17.0
feat(guardrails): guard chains and ready-made detectors by @DEENUU1 in https://github.com/pydantic/pydantic-ai-harness/pull/478
system_reminders): cache-safe SystemReminders capability by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/414Full Changelog: https://github.com/pydantic/pydantic-ai-harness/compare/v0.15.0...v0.16.0
system_reminders): cache-safe SystemReminders capability by @dsfaccini in #414Full Changelog: v0.15.0...v0.16.0
feat: add StackOne capability for linked account actions by @adtyavrdhn in https://github.com/pydantic/pydantic-ai-harness/pull/479
agent_name storage key by @sevakva in https://github.com/pydantic/pydantic-ai-harness/pull/449ToolGuard for tool arguments and tool results by @DEENUU1 in https://github.com/pydantic/pydantic-ai-harness/pull/470agent_docs): require an issue comment when docs or code link it; retry the dependency gate by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/515max_output_chars cap accounting by @adtyavrdhn in https://github.com/pydantic/pydantic-ai-harness/pull/507Shell allowlist with the default denylist by @adtyavrdhn in https://github.com/pydantic/pydantic-ai-harness/pull/508run_code again (revealed_tool_names) by @DouweM in https://github.com/pydantic/pydantic-ai-harness/pull/517Full Changelog: https://github.com/pydantic/pydantic-ai-harness/compare/v0.14.0...v0.15.0
agent_name storage key by @sevakva in #449ToolGuard for tool arguments and tool results by @DEENUU1 in #470agent_docs): require an issue comment when docs or code link it; retry the dependency gate by @dsfaccini in #515max_output_chars cap accounting by @adtyavrdhn in #507Shell allowlist with the default denylist by @adtyavrdhn in #508run_code again (revealed_tool_names) by @DouweM in #517Full Changelog: v0.14.0...v0.15.0
fix(ci): let dependency approval revoke a label and still report by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/495
Full Changelog: https://github.com/pydantic/pydantic-ai-harness/compare/v0.13.0...v0.14.0
subagents: per-delegation model selection via an opt-in model menu by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/451
step_persistence: opt-in max_snapshots_per_run to bound snapshot growth by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/442docs/nav.json parity so pages cannot ship orphaned from the site nav by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/491conversation_search): BM25 search over the history StepPersistence stores by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/413Full Changelog: https://github.com/pydantic/pydantic-ai-harness/compare/v0.12.0...v0.13.0
Add Agent Skills as deferred capabilities by @adtyavrdhn in https://github.com/pydantic/pydantic-ai-harness/pull/396
Full Changelog: https://github.com/pydantic/pydantic-ai-harness/compare/v0.11.0...v0.12.0
Exclude agent state from typechecking by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/435
Full Changelog: https://github.com/pydantic/pydantic-ai-harness/compare/v0.10.0...v0.11.0
feat: add ModalSandbox for isolated cloud sandboxes by @strawgate in https://github.com/pydantic/pydantic-ai-harness/pull/269
ClampOversizedMessages clamping typed ToolCallPart subclasses (ToolSearchCallPart, LoadCapabilityCallPart) by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/411Memory capabilities coexist by @sevakva in https://github.com/pydantic/pydantic-ai-harness/pull/409release job silently skipping on tag pushes by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/429release skipping on tag pushes: opt out of skipped-ancestor propagation by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/430Full Changelog: https://github.com/pydantic/pydantic-ai-harness/compare/v0.9.0...v0.10.0
> This release raises the pydantic-ai-slim floor to 2.14.1.
[!IMPORTANT] This release raises the
pydantic-ai-slimfloor to 2.14.1.
cache_stability): observational CacheStabilityMonitor capability by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/327web_search text_summary mode and deferred ExaAgent capability by @ryahern in https://github.com/pydantic/pydantic-ai-harness/pull/386ClearToolResults corrupting typed ToolReturnPart subclasses (ToolSearchReturnPart) by @dsfaccini in https://github.com/pydantic/pydantic-ai-harness/pull/383Full Changelog: https://github.com/pydantic/pydantic-ai-harness/compare/v0.8.0...v0.9.0
Add elicitation callback support to MCP servers by @yamanahlawat in https://github.com/pydantic/pydantic-ai/pull/2373
message_history parameter to agent.to_cli() by @ryx2 in https://github.com/pydantic/pydantic-ai/pull/2674FileUrl.media_type by @mwildehahn in https://github.com/pydantic/pydantic-ai/pull/2677Full Changelog: https://github.com/pydantic/pydantic-ai/compare/v0.7.6...v0.8.0
Add new OpenAI GPT-5 models by @medaminezghal in https://github.com/pydantic/pydantic-ai/pull/2503
"openai-responses" model inference string by @Kludex in https://github.com/pydantic/pydantic-ai/pull/2528UsageLimits.count_tokens_before_request using Gemini count_tokens API by @kauabh in https://github.com/pydantic/pydantic-ai/pull/2137FallbackModel to respect each model's model settings by @jerry-reevo in https://github.com/pydantic/pydantic-ai/pull/2540Full Changelog: https://github.com/pydantic/pydantic-ai/compare/v0.7.0...v0.7.1
Let Agent be run in a Temporal workflow by moving model requests, tool calls, and MCP to Temporal activities by @DouweM in https://github.com/pydantic
event_stream_handler to agent and run methods by @DouweM in https://github.com/pydantic/pydantic-ai/pull/2458AbstractAgent and WrapperAgent by @DouweM in https://github.com/pydantic/pydantic-ai/pull/2458Agent.override(tools=...) by @DouweM in https://github.com/pydantic/pydantic-ai/pull/2458Full Changelog: https://github.com/pydantic/pydantic-ai/compare/v0.6.2...v0.7.0
BREAKING CHANGE: Remove next() method from Graph by @Kludex in https://github.com/pydantic/pydantic-ai/pull/2440
next() method from Graph by @Kludex in https://github.com/pydantic/pydantic-ai/pull/2440data from FinalResult by @Kludex in https://github.com/pydantic/pydantic-ai/pull/2443get_data and validate_structured_result from StreamedRunResult by @Kludex in https://github.com/pydantic/pydantic-ai/pull/2445griffe_warnings_deprecated by @Kludex in https://github.com/pydantic/pydantic-ai/pull/2444format_as_xml module by @Kludex in https://github.com/pydantic/pydantic-ai/pull/2446result_type parameter and similar from Agent by @Kludex in https://github.com/pydantic/pydantic-ai/pull/2441GoogleGLAProvider and GoogleVertexProvider by @Kludex in https://github.com/pydantic/pydantic-ai/pull/2450Full Changelog: https://github.com/pydantic/pydantic-ai/compare/v0.5.1...v0.6.0
Breaking Change: The EvaluationReport.print and EvaluationReport.console_table methods now require most arguments be passed by keyword. by @dmontagu i…
EvaluationReport.print and EvaluationReport.console_table methods now require most arguments be passed by keyword. by @dmontagu in https://github.com/pydantic/pydantic-ai/pull/2163source field of an EvaluationResult is now of type EvaluatorSpec rather than the actual source Evaluator instance, to help with serialization/deserialization by @dmontagu in https://github.com/pydantic/pydantic-ai/pull/2388BaseModels use OpenAI strict JSON mode by defaulting to additionalProperties=False by @DouweM in https://github.com/pydantic/pydantic-ai/pull/2419test_download_item_no_content_type test failing on macOS by @strawgate in https://github.com/pydantic/pydantic-ai/pull/2404install.md by @Kludex in https://github.com/pydantic/pydantic-ai/pull/2412Full Changelog: https://github.com/pydantic/pydantic-ai/compare/v0.4.11...v0.5.0
BREAKING CHANGE: Make EvaluationReport and ReportCase into generic dataclasses by @dmontagu in https://github.com/pydantic/pydantic-ai/pull/1799
EvaluationReport and ReportCase into generic dataclasses by @dmontagu in https://github.com/pydantic/pydantic-ai/pull/1799ToolDefinition.description optional and fix Bedrock description handling by @dmontagu in https://github.com/pydantic/pydantic-ai/pull/1507Full Changelog: https://github.com/pydantic/pydantic-ai/compare/v0.3.7...v0.4.0
See #1142 — Adds support for thinking parts.
See #1142 — Adds support for thinking parts.
We now convert the thinking blocks ("<think>..."</think>") in provider specific text parts to
PydanticAI ThinkingParts. Also, as part of this release, we made the choice to not send back the
ThinkingParts to the provider - the idea is to save costs on behalf of the user. In the future, we
intend to add a setting to customize this behavior.
Full Changelog: https://github.com/pydantic/pydantic-ai/compare/v0.2.20...v0.3.0
Add AWS Profile Feature by @prescod in https://github.com/pydantic/pydantic-ai/pull/1634
~/.pydantic-ai by @Kludex in https://github.com/pydantic/pydantic-ai/pull/1701BinaryContent information by @Kludex in https://github.com/pydantic/pydantic-ai/pull/1603Full Changelog: https://github.com/pydantic/pydantic-ai/compare/v0.2.0...v0.2.1
See #1647 — usage makes sense as part of ModelResponse, and could be really useful in "messages" (really a sequence of requests and response). In this
See #1647 — usage makes sense as part of ModelResponse, and could be really useful in "messages" (really a sequence of requests and response). In this PR:
usage to ModelResponse (field has a default factory of Usage() so it'll work to load data that doesn't have usage)Model.request to just ModelResponse instead of tuple[ModelResponse, Usage]ToolCallPartDelta.tool_call_id when known from previous delta by @DouweM in https://github.com/pydantic/pydantic-ai/pull/1694Full Changelog: https://github.com/pydantic/pydantic-ai/compare/v0.1.11...v0.2.0
Fix max length handling by @dmontagu in https://github.com/pydantic/pydantic-ai/pull/1510
BinaryContent by decoding base64 by @DouweM in https://github.com/pydantic/pydantic-ai/pull/1513StdioServerParameters.cwd param by @timesler in https://github.com/pydantic/pydantic-ai/pull/1514Full Changelog: https://github.com/pydantic/pydantic-ai/compare/v0.1.1...v0.1.2
fix: Allow empty tool call args by @Wh1isper in https://github.com/pydantic/pydantic-ai/pull/1407
UserPromptNode by @Kludex in https://github.com/pydantic/pydantic-ai/pull/1503additionalProperties by gemini by @Kludex in https://github.com/pydantic/pydantic-ai/pull/1506Full Changelog: https://github.com/pydantic/pydantic-ai/compare/v0.1.0...v0.1.1