OpenAI Agents SDK changelog: what's new each month
Every stable OpenAI Agents SDK release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 11 months covered; the current month updates daily.
October 2026
3 releases: 0.23.0 → 0.24.0
October brought refinements to sandbox configuration, session management, and tool handling across three releases. The month focused on expanding customization options and improving reliability in core agent operations.
Highlights
- MCP tools now expose annotations and support run-wide exception formatting for better error handling.
- Sandbox operations gained opt-in Docker removal protection and configurable memory consolidation.
- Session encryption and conversation writes received new budget and batching controls to optimize performance.
- Sensitive trace capture now respects model metadata to improve privacy handling.
New
- Configurable MCP listing page limits.
- Opt-in Docker removal protection for sandbox environments.
- Encrypted history scan budget for sessions.
- MCP annotations exposed on converted tools.
- Run-wide tool exception formatting support.
- Configurable memory consolidation turns for sandbox operations.
Improved
- Conversation session writes now batched within API limits for efficiency.
- Automatic compaction deferred for pending approvals to prevent conflicts.
- Sensitive trace capture now respects model metadata classification.
- Cleanup operations no longer await during coroutine closure, reducing shutdown delays.
Fixed
- Release tests restored and required readiness check refreshed in 0.23.1.
- Test-loop initialization and required-check ownership corrected.
September 2026
3 releases: 0.22.1 → 0.22.3
September brought refinements to tool execution, sandboxing, and guardrail customization across the SDK. The month focused on stability improvements and expanding capability options for web search, voice, and image generation.
Highlights
- Web search tools now return image results alongside text.
- Customize the messages guardrails display when blocking content.
- Configure Unix-local sandbox isolation and add labels to Docker containers.
- Improved stability in session management, symlink handling, and concurrent operations.
New
- Server-wide guardrails for MCP tools.
- Configurable Unix-local environment isolation in sandboxes.
- Labels support on Docker sandbox containers.
- Streamed transcription options exposed in voice handling.
- Support for current image generation tool options.
Improved
- Web search tools now surface image results in addition to text.
- Guardrail blocked messages can be customized.
- Tool argument validation aligned with conditional approvals.
- POSIX command paths maintained on Windows hosts running sandboxes.
- Session compaction response chain handling after pop operations.
Fixed
- Detached aggregated request usage entries to prevent tracking issues.
- Tool-not-found output now delivered correctly on server-managed resume.
- Symlink race conditions prevented in UnixLocal file API.
- Concurrent async SQLite startup now handles correctly.
- Session limit no longer incorrectly forwarded as Conversations page size.
- Subprocesses reaped on early stream close to prevent leaks.
August 2026
7 releases: 0.19.2 → 0.22.0
August brought significant improvements to agent reliability, guardrail enforcement, and testing capabilities. The month also introduced major feature additions including a new default model, durable input staging, and comprehensive testing utilities.
- Applications using custom MCP HTTP authentication or client factories must use HTTP types from the installed MCP major version, or pin mcp<2.
- OpenAI provider now rejects conflicting options when receiving an explicit client.
- RunState interruption snapshots and max-turn finalization behavior hardened; may affect existing custom workflows.
- Non-streaming Responses with terminal status failed or incomplete now raise ModelBehaviorError.
Highlights
- New default model gpt-5.6-luna with support for explicit overrides and environment variable configuration.
- Added agents.testing utilities for deterministic testing without making provider requests.
- RunState.addinput() enables staging durable user input before resumed model calls with full guardrail support.
- Upgraded to OpenAI Python SDK 3.x with HTTPX2 compatibility.
- MCP now supports both Python SDK v1 and v2 across multiple transport protocols.
New
- Durable pending input staging through RunState.addinput() with guardrail and persistence support.
- Testing utilities for Agent, Sandbox, Realtime, and Voice workflows for deterministic testing.
- Model call timeouts for improved control over execution duration.
- Run-scoped sandbox working directories for better isolation.
- Docker sandbox networking disable option.
- Modal sandbox resource options.
- Realtime transcription support for gpt-live-transcribe, gpt-transcribe, and gpt-realtime-whisper.
- Sandbox Runloop existing-secret support.
Improved
- Guardrail results now reported for streamed runs and tool executions.
- Max-turns handler output persists to session.
- Tool name collision resolution made consistent across agent and resumed runs.
- Output guardrails deferred until after non-stream session saves.
- Sandbox mount validation with explicit credential-exposure acknowledgements.
- MCP lifecycle snapshot isolation and configurable retry backoff ceilings.
- Voice validation for invalid channels, frame rates, and audio rates.
- Agent graphs now expand agents registered through handoff().
Fixed
- Input and output guardrail results now properly reported when tripwires abort runs.
- Tool guardrail results preserved and honored during streamed resume operations.
- Closed default event loop in runsync replaced.
- Non-streaming content-filter refusals now surfaced correctly.
- Sibling work cancelled after concurrent failures.
- Terminal function-tool output rejected by agent output guardrails redacted from replay state.
July 2026
7 releases: 0.17.8 → 0.19.1
July brought significant expansions to the SDK's capabilities, including programmatic tool calling, multi-agent support, and improved model compatibility. The month focused on reliability improvements, better error handling, and support for newer model features.
Highlights
- Programmatic Tool Calling support lets models generate JavaScript to coordinate eligible tools with per-tool caller restrictions.
- Hosted multi-agent beta support enables coordinating multiple agents in a single deployment.
- GPT-5.6 model support with request controls and new default realtime model gpt-realtime-2.1.
- Configuration now accepts typed settings objects or dictionaries consistently across all components with validation.
- Hardened logging across the SDK redacts sensitive payloads while preserving debugging context.
New
- ProgrammaticToolCallingTool for supported OpenAI response models to generate JavaScript tool coordination.
- Public agents.decorators module with shorter @tool alias for defining function tools.
- Async callable object support for function tools.
- VercelCloudBucketMountStrategy for mounted sessions with intentional workspace persistence exclusions.
- Configurable task and turn tracing spans.
- Response usage tracking in realtime session context.
- Native host paths support in sandbox path grants.
- Unicode storage option for SQLAlchemy sessions.
Improved
- Session history is now preserved across model retries.
- Chat Completions streams close properly on early exit.
- Realtime session cleanup is now deterministic with monotonic playback timing.
- Nested tool state is preserved during restoration.
- Improved AnyLLM and LiteLLM compatibility with better cache-write support across Python versions.
- Streamed logprobs accumulation optimized to avoid O(n²) performance overhead.
- Error messages use UserError for conflicting provider arguments instead of assertions.
- Content-filter refusals from LiteLLM are now surfaced instead of empty turns.
Fixed
- Realtime validation failure logs are redacted to remove sensitive information.
- Visualization preserves fillcolor on handoff nodes by properly merging style attributes.
- Invalid final output recovery handler prevents crashes on malformed outputs.
- Quotes and backslashes in DOT names are properly escaped for visualization.
- List content in AdvancedSQLiteSession browsing helpers is now handled correctly.
- WebSocket server errors occurring before response starts are retried.
June 2026
3 releases: 0.17.5 → 0.17.7
June brought improvements to tool handling, guardrails, and streaming capabilities across three releases. The SDK focused on reliability fixes and new configuration options for production use cases.
Highlights
- Added pre-approval tool input guardrails to control what agents can execute.
- Introduced buffered Chat Completions tool-call streaming for better performance.
- Exposed configurable websocket maxsize limit for production deployments.
- Added SDK-only custom data support for tool outputs.
- Fixed critical sandbox and session atomicity issues.
New
- Pre-approval tool input guardrails to validate tool calls before execution.
- SDK-only custom data for tool outputs.
- Configurable websocket maxsize limit.
- Buffered Chat Completions tool-call streaming.
- MongoDB session example in documentation.
Improved
- Sandbox error retryability is now exposed for better error handling.
- Tool output types properly reflect object structure.
- Reduced sandbox sink buffering and spool timeouts for faster operations.
- Suppressed spurious warnings about handoff whitespace in tool names.
- Made AdvancedSQLiteSession addItems operation atomic for data consistency.
Fixed
- Enforce documented strict JSON-compatible contract for tool inputs.
- Clean up branch-only messages when deleting branches.
- Cancel sibling guardrail tasks when one raises an error.
- Skip needs-approval checker when status is already resolved in collect-runs-by-approval.
- Validate special permission bits by position in file operations.
- Fixed type casting for tool-end hook results.
May 2026
11 releases: 0.15.0 → 0.17.4
May brought multiple stability and security improvements across the SDK, with updates to model defaults, session handling, and sandbox constraints. The month saw steady refinements to tool handling, tracing, and real-time audio support.
- Default model changes to a newer mini model when model is unset; explicitly set a model if you need previous behavior.
- Default real-time sessions now use gpt-realtime-2 instead of previous default.
Highlights
- Default model switched to a newer mini model for improved performance when no model is specified.
- Added opt-in server-prefixed MCP tool names and runtime function tool concurrency configuration.
- Default real-time sessions now use gpt-realtime-2 with support for custom voice objects.
- Multiple fixes to session state management, tool output stability, and sandbox security.
- Improved error handling and redaction for logs, plus better support for disabled tool logging.
New
- Expose Responses WebSocket keepalive options for connection tuning.
- Add context management model setting to control agent behavior.
- Allow disabling max turns with None to run agents without turn limits.
- Add opt-in server-prefixed MCP tool names for tool namespace management.
- Add runtime function tool concurrency configuration.
- Support Realtime custom voice objects for audio interactions.
- Add opt-in recovery for missing function tools.
- Export additional tracing-related functions and types from agents module.
Improved
- Restore UnixLocal PTY terminal signal defaults and SIGINT handling.
- Improve quickstart documentation for Windows OS users.
- Stabilize chat completions tool call output indexes for consistent ordering.
- Better error messages from sandbox providers with detailed error information.
- Redact function tool trace span errors and invalid JSON errors when appropriate.
- Make tracing shutdown best-effort on process exit to prevent hang-ups.
- Make Permissions hashable to match User and Group types.
- Preserve GitRepo root subpath aliases and handle empty subpaths correctly.
Fixed
- Surface model refusals during run resolution so they are visible to users.
- Avoid replaying assistant conversation item IDs for OpenAIConversationsSession.
- Reject failed responses stream terminals and corrupt session state updates.
- Fix OpenAI Conversations reasoning persistence across sessions.
- Constrain local sandbox artifact sources to base directory for security.
- Limit sandbox archive extraction and validate git repository subpaths.
April 2026
12 releases: 0.13.4 → 0.14.8
April brought a major sandbox feature release alongside steady improvements to session handling, tracing, and cross-platform compatibility. The SDK now includes comprehensive sandboxing capabilities with multiple client implementations and support for workflow snapshots.
Highlights
- New SandboxAgent with built-in capabilities for shell access, filesystem editing, image inspection, and memory management.
- Multiple sandbox client implementations available: local Unix and Docker clients, plus hosted options for Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel.
- Workspace snapshots and serialized session state let you reconnect to existing sandbox work or seed fresh sandboxes from saved contents.
- MongoDB session backend added as an extension option.
- Windows support for sandboxes with path normalization and native CI testing.
New
- SandboxAgent with sandbox-specific defaults, instructions, and capabilities.
- Manifest contract for specifying files, directories, Git repos, environment, users, groups, and mounts.
- SandboxRunConfig for per-run sandbox wiring including client creation and manifest overrides.
- Workspace snapshots for sandbox state persistence.
- Public flushtraces API.
- Callable approval policies for local MCP servers.
- BoxMount support for sandboxes.
- Modal sandbox idle timeout option.
Improved
- AnyLLM response handling and reasoning extraction for iterable objects.
- SQLite session writes now use shared file locks for serialization.
- SQLiteSession and SQLAlchemySession hardened against transient locks.
- Trace preview no longer recursively truncates.
- Tool origin metadata persisted in run items.
- Phase 2 memory consolidation turn limits increased.
- Computer driver modifier key compatibility preserved.
- Convenience properties (toolname, callid) added to tool items.
Fixed
- AnyLLM response sanitization before validation.
- Sandbox path normalization and Windows compatibility issues.
- MCP re-export import error preservation.
- HITL resume tool outputs now served correctly.
- Streamed terminal output backfilling.
- Symlinked LocalFile sources now rejected for security.
March 2026
15 releases: 0.10.3 → 0.13.3
March brought significant enhancements to the OpenAI Agents SDK, including a major model update to gpt-realtime-1.5, expanded MCP capabilities, and improved reliability across runtime and session management. The month saw steady releases with a focus on edge case fixes and provider compatibility improvements.
Highlights
- Realtime agents now default to gpt-realtime-1.5 model for improved performance without requiring configuration changes.
- MCP servers now support resource listing and resumable streaming HTTP sessions with session IDs for connection recovery.
- New any-LLM model support enables responses-compatible routing across multiple language model providers.
- Chat Completions integrations can now replay reasoning content, improving tool-call continuity for provider-specific adapters.
- Tool search support added via the Responses API for enhanced agent capability discovery.
New
- Responses API tool search support for agents.
- Opt-in model retry policies with configurable jitter and delay caps.
- Explicit MultiProvider prefix modes for better provider handling.
- Any-LLM model support with responses-compatible routing.
- External web access option for WebSearchTool.
- Auth and httpxclientfactory exposure in MCP SSE and StreamableHttp parameters.
- Reasoning-content replay option for Chat Completions integrations via shouldreplayreasoningcontent.
- MCP resource listing and resource template capabilities on MCPServer.
Improved
- Immutable agent tool invocation metadata now exposed on run results.
- RunResult now includes toolcontext accessor for agent tools.
- BFS queue and voice task ordering optimized using deque for better performance.
- MCPServerStreamableHttp now exposes sessionid for resumable streaming HTTP sessions.
- Structured realtime tool outputs now serialized as JSON.
- Reasoning summaries now finish before text deltas are sent.
- Documentation expanded across guides for tool search and GA computer tool migration.
- Improved documentation for Dapr sessions, advanced model settings, and realtime transport boundaries.
Fixed
- Isolated parallel function tool failures to prevent cascade failures.
- MCP errors now returned as structured error results instead of crashing agent runs.
- ComputerTool migrated to GA computer tool with optional preview metadata.
- Concurrent first writes in SQLAlchemySession now handled correctly.
- MCP tool cancellations and ClosedResourceError now properly handled and retried.
- Orphan hosted shell calls dropped before multi-turn replay to prevent state corruption.
February 2026
12 releases: 0.8.0 → 0.10.2
February brought significant expansions to agent capabilities and reliability, with major features for human-in-the-loop workflows, structured tool inputs, and configurable timeouts. The month saw rapid iteration across eleven releases, culminating in version 0.10.2.
- Python 3.9 support dropped; Python 3.10 or later required
- Local MCP tool failure behavior changed to return model-visible errors by default; set timeout behavior explicitly if you need fail-fast semantics
- RunState schema updated to version 1.1 with explicit backward-read policy
Highlights
- Human-in-the-loop support enables pausing agent execution for human approval or intervention.
- Structured agent tool input support provides better type safety and validation for tool parameters.
- Configurable timeout handling for function tools lets you define how timeouts are handled per tool.
- Container shell and skills support expands agent capabilities for system-level operations.
- Responses websocket model enables real-time bidirectional communication patterns.
New
- Human-in-the-loop support for agent workflows
- Structured tool input support using Pydantic annotations
- Max turns run error handlers for controlling agent iteration limits
- Session customization parameters
- MCP tool metadata resolver support
- Image responses from MCP servers
- Container shell and skills support
- Responses websocket model for streaming communication
Improved
- Function tools now execute on worker threads instead of the event loop thread
- Local MCP tool failures are configurable with model-visible error output as default
- Tool context now includes agent information in tool calls
- Pydantic Field annotations are supported for tool arguments
- Run-context thread reuse for code execution tools
- Max turns limit now applies in REPL environments
- Tool timeout behavior is now configurable with options for error handling
- Trace metadata propagation improved across processors and spans
Fixed
- Streamed run-again tool items now persist to session correctly
- Handoff target resolution compatibility preserved in run-state agent maps
- Nested agent-tool HITL state persists across RunState JSON round-trips
- Tracing initialization hazards that could occur on fork eliminated
- Oversized tracing span payloads are now sanitized to prevent failures
- Resumed traces reattach without duplicate trace starts
January 2026
6 releases: 0.6.5 → 0.7.0
January brought refinements to conversation management, model compatibility, and tool handling, with significant improvements to session storage and reasoning configurations. The month culminated in version 0.7.0 with breaking changes to default reasoning behavior and session input handling.
- Default reasoning.effort for gpt-5.1 and gpt-5.2 changed from 'low' to 'none'; explicitly set to 'low' if previous behavior is required.
- Session input callback (sessioninputcallback) is now optional when using a sessions store; default behavior appends new input to session history.
- Nested handoffs moved to opt-in feature and are no longer enabled by default.
Highlights
- Auto-compaction for long conversations now available to reduce context size and improve performance.
- Async session storage support via AsyncSQLiteSession for non-blocking persistence.
- Experimental Codex extension and tool for enhanced code-related capabilities.
- Nested handoffs moved to opt-in to simplify agent workflows by default.
- Session input auto-append now default behavior, removing the requirement for manual input callbacks.
New
- Per-run tracing API key support for flexible trace authentication.
- AgentHookContext with turn input parameter for agent hooks.
- Tool guardrails in functiontool decorator arguments.
- Realtime audio mapping support and SIP session payload testing.
- Input-based responses compaction with store-aware auto mode.
- MCPServerManager for safely managing MCP server lifecycle.
- WebSocket custom options for OpenAIRealtimeWebSocketModel.
- Experimental Codex extension tool for code operations.
Improved
- Tool execution exceptions now converted to runtime errors for consistent error handling.
- Realtime text mode now handles outputtext type correctly.
- PCM duration calculation improved and VAD truncation handled better.
- Thinking tokens now attached properly even without tool calls.
- OPENAIAGENTSDISABLETRACING environment variable loading improved.
- Tracing shutdown made safer against closed handlers.
- Exception handling enhanced in MCP server initialization and cleanup.
- Shell tool outputs now enforced with maxoutputlength limits.
Fixed
- Streaming blocking and inability to be cancelled resolved.
- Invalid tool arguments JSON no longer crashes the system.
- Nested handoff inputs deduplication fixed.
- Tool message ordering corrected for Gemini models.
- DeepSeek reasoning content now included in tool call messages.
- Compaction deferred when local tool outputs are present.
December 2025
1 release: 0.6.4
December brought refinements to agent tooling and streaming capabilities, along with fixes for function schema handling and token usage tracking.
Highlights
- Agents can now use onstream as a tool for handling streamed outputs.
- Computer resources can now be created and disposed per agent run for better resource management.
- Fixed missing strict JSON schema validation when using Chat Completions API.
- Improved null handling for token usage details.
New
- onstream tool support for agents
Improved
- Computer resource lifecycle can now be managed independently per agent run
- Null handling enhanced for input and output token details in Usage tracking
- failureerrorfunction now exposed in Agent.astool for better error handling
Fixed
- FunctionTool.strictjsonschema no longer missing when using Chat Completions API
- Null handling in Usage class for token detail fields
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. OpenAI Agents SDK: pricing, features and alternatives · All changelogs