v1.21.3

Bug Fixes 🐛

• Fix E2BExecutor by pinning e2b-code-interpreter to < 2 by @albertvillanova in #1733

Full Changelog: v1.21.2...v1.21.3

v1.21.2

Bug Fixes 🐛

• Pin telemetry extra >=0.1.15 to use new TokenUsage structure by @albertvillanova in #1699

Full Changelog: v1.21.1...v1.21.2

v1.21.1

Bug Fixes 🐛

• Fix error for GPT-5 not supporting the stop parameter by @Xmader in #1662

Full Changelog: v1.21.0...v1.21.1

v1.21.0

Enhancements ✨

• Tool & Model Improvements

  • Add Tool to_code_prompt and to_tool_calling_prompt by @albertvillanova in #1599
  • Improve function body extraction in tools by @HairlessVillager in #1627
  • Support passing model_kwargs to TransformersModel by @albertvillanova in #1608
  • Support dict messages as direct input to models by @albertvillanova in #1570

• Executor Improvements

  • Enhance LocalPythonExecutor security by blocking dunder calls by @nnfrog in #1551
  • Fix continue semantics of LocalPythonExecutorby @adaminsky in #1645
  • Fix GeneratorExp in LocalPythonExecutor to return Generator instead of list by @albertvillanova in #1650

• Gradio & UI

  • Reset agent memory on GradioUI clear button by @albertvillanova in #1611
  • Use gr.Number for mapping integer and number type components in launch_gradio_demo by @karenwky in #1320

Bug Fixes 🐛

• Fix: Prevent empty additional_args from being added to agent prompt by @jokokojote in #1566
• Fix SyntaxError in generated Gradio app due to 'class' attribute by @albertvillanova in #1615
• Fix serialization inconsistency in MultiStepAgent.to_dict/from_dict for managed_agents by @hazrmard in #1577
• Fix a kwargs error in LocalPythonExecutor by @aymeric-roucher in #1583
• Fix structured outputs by @MicrosoftJames in #1642
• Fix bug where AmazonBedrockServerModel crashes in thinking mode by @benoriol in #1632
• Fix: Remove duplicate line in tool_calling_agent prompt by @yusuf-eren in #1636
• Fix: RemotePythonExecutor should not send variables if none exist by @thomasjpfan in #1605
• Disable parallel tool calls for final answer by @aymeric-roucher in #1539

Documentation 📚

• Improve Installation docs with uv option by @albertvillanova in #1550
• Fix and improve MCP server security warnings in docs by @albertvillanova in #1549
• DOC: Fix image link for docs (secure_code_execution.md) by @FacerAin in #1582
• Update title of page "What are agents" by @aymeric-roucher in #1546
• Enable multilingual docs by @albertvillanova in #1633
• 🌐 [i18n-KO] Translated index.md to Korean by @FacerAin in #1581
• 🌐 [i18n-KO] Translated text_to_sql.md to Korean by @nsbg in #1603

CI / Dependency Updates 🧪

• Pin mcpadapt to support image/audio content by @albertvillanova in #1558
• Temporarily skip wrongly implemented tests by @albertvillanova in #1571
• CI hotfix: Pin mlx < 0.26.5 by @albertvillanova in #1586
• Support mlx-0.26.5 by @albertvillanova in #1589
• CI hotfix: Pin transformers < 4.54.0 by @albertvillanova in #1620
• Support transformers 4.54.0 by @albertvillanova in #1625
• Fix CI PytestUnknownMarkWarning: Unknown pytest.mark.timeout by @albertvillanova in #1630
• Fix CI quality: remove trailing whitespace by @albertvillanova in #1617

Deprecations & Cleanups 🧹

• Replace deprecated ast.Num and ast.Str with ast.Constant in tool validation by @WilliamRabuel in #1594
• Remove deprecated grammar parameter by @albertvillanova in #1651
• Remove deprecated token count attributes by @albertvillanova in #1652
• Remove deprecated agent logs attribute by @albertvillanova in #1653
• Remove mention of deprecated ManagedAgent from docs by @albertvillanova in #1655
• Remove deprecated default sse transport by @albertvillanova in #1656
• Replace duckduckgo_search with ddgs to address package rename warning by @WilliamRabuel in #1593
• Remove unused is_same_item function by @albertvillanova in #1631
• Remove unnecessary import by @rahul-1729 in #1639
• Simplify _is_package_available by @albertvillanova in #1641

Miscellaneous 🔧

• Add support for Amazon Bedrock API Keys by @colesmcintosh in #1612
• Factorize create_agent_gradio_app_template function by @albertvillanova in #1616
• Fix: Allow passing integers for floats by @aymeric-roucher in #1623
• Fix Tool.forward to raise NotImplementedError instead of returning by @HairlessVillager in #1628
• Make validate_tool_arguments raise instead of returning error string by @albertvillanova in #1634
• Remove mention of python in additional args by @aymeric-roucher in #1584
• ToolCallingAgent: remove unnecessary addition to model_output by @aymeric-roucher in #1622
• Minor fix in toolCallingAgentPrompt by @rahul-1729 in #1574

Maintenance 🏗️

• Bump dev version: v1.21.0.dev0 by @albertvillanova in #1544
• Remove empty issue template by @albertvillanova in #1562
• Improve issue templates by @albertvillanova in #1563
• Add issue template for documentation improvements by @albertvillanova in #1564
• Add uv option to CONTRIBUTING by @albertvillanova in #1561

New Contributors

• @jokokojote made their first contribution in #1566
• @rahul-1729 made their first contribution in #1574
• @WilliamRabuel made their first contribution in #1594
• @FacerAin made their first contribution in #1582
• @thomasjpfan made their first contribution in #1605
• @hazrmard made their first contribution in #1577
• @HairlessVillager made their first contribution in #1628
• @benoriol made their first contribution in #1632
• @yusuf-eren made their first contribution in #1636
• @nnfrog made their first contribution in #1551
• @nsbg made their first contribution in #1603
• @karenwky made their first contribution in #1320
• @MicrosoftJames made their first contribution in #1642
• @adaminsky made their first contribution in #1645

Full Changelog: v1.20.0...v1.21.0

v1.20.0

New Features ✨

• Remote Execution:

  • Implement remote Python WasmExecutor by @albertvillanova in #1261

• Callbacks:

  • Support callbacks after planning step by @albertvillanova in #1445
  • Support multiple callbacks per step type in step_callbacks dict by @albertvillanova in #1517

Enhancements 🛠️

• Models:

  • Implement rate limiting in all API models by @albertvillanova in #1516

• Tools:

  • Implement exception mechanism for final answer in remote executors by @tobiasofsn in #1429
  • Set default api_key_name in ApiWebSearchTool by @albertvillanova in #1507
  • Implement rate limit mechanism in ApiWebSearchTool by @albertvillanova in #1508
  • Implement rate limit mechanism in DuckDuckGoSearchTool by @albertvillanova in #1515
  • Enable image output for Tool.from_space by @aymeric-roucher in #1510
  • Add adapter_kwargs parameter to MCPClient by @LucasDltg in #1528

• Validation of Tool Arguments:

  • Support multiple types in tool arguments validation by @duguyue100 in #1505
  • Fix input type validation when input schema has multiple types by @duguyue100 in #1530
  • Fix optional args toolcalling by @aymeric-roucher in #1538

• Agent Output Formatting:

  • Implement CodeOutput as analog to ToolOutput by @albertvillanova in #1496
  • Allow markdown or custom formatting for code blocks by @aymeric-roucher in #1493

• Other:

  • Add ApiWebSearchTool to all list by @albertvillanova in #1492
  • Enforced type annotations in ChatMessage via MessageRole enum by @daavoo in #1479
  • Check authorized imports in LocalPythonExecutor constructor by @jank in #265

Bug Fixes 🐛

• Fix KeyError when agent reaches max steps with image input by @robin-ede in #1525
• Fix when response.usage is None for some OpenRouter models by @sentienthouseplant in #1402
• Fix incorrect token counting in streaming TransformersModel by @albertvillanova in #1503
• Minor fix nested quotes that prevent execution by @vinoo-ganesh-keru in #1497
• Fix ChatMessage attribute access after replacing dict with dataclass by @albertvillanova in #1534
• Fix access of content field in ChatMessage by @mdr223 in #1533
• Fix parallel final answers handling by @aymeric-roucher in #1482

Documentation Improvements 📚

• Examples & Tutorials:

  • Add complete Plan Customization example by @albertvillanova in #1506
  • Add Async App with Agent example by @albertvillanova in #1509
  • Improve Plan Customization example with Human-in-the-Loop by @albertvillanova in #1513

• API References:

  • Fix docstring of WikipediaSearchTool by @albertvillanova in #1521
  • Explain rate_limit in tools docstrings by @albertvillanova in #1519
  • Fix Examples in AmazonBedrockServerModel docstring by @albertvillanova in #1522
  • Split Built-in Tools from Tools in Reference docs by @albertvillanova in #1520

• Other:

  • Update Langfuse tracing documentation to SDK v3 by @jannikmaierhoefer in #1447
  • Fix use of context manager agent in docs by @albertvillanova in #1502

Maintenance 🏗️

• Fix CI LiteLLM test_call_different_providers_without_key by @albertvillanova in #1527
• Fix test_send_tools_with_default_wikipedia_search_tool by @albertvillanova in #1494
• Fix DockerExecutor tests with final_answer by calling send_tools by @albertvillanova in #1495
• Bump dev version: v1.20.0.dev0 by @albertvillanova in #1477

New Contributors

• @sentienthouseplant made their first contribution in #1402
• @daavoo made their first contribution in #1479
• @duguyue100 made their first contribution in #1505
• @vinoo-ganesh-keru made their first contribution in #1497
• @robin-ede made their first contribution in #1525
• @LucasDltg made their first contribution in #1528
• @mdr223 made their first contribution in #1533

Full Changelog: v1.19.0...v1.20.0

v1.19.0

Enhancements 🛠️

• Agent Upgrades:

  • Support managed agents in ToolCallingAgent by @albertvillanova in #1456
  • Support context managers for agent cleanup by @tobiasofsn in #1422
  • Change code tags to xml by @aymeric-roucher in #1442

• UI Improvements:

  • Support reset_agent_memory in GradioUI by @JakeBx in #1420

• Streaming Refactor:

  • Transfer aggregation of streaming events off the Model class by @aymeric-roucher in #1449

• Agent Output Tracking:

  • Store CodeAgent code outputs in ActionStep by @aymeric-roucher in #1463

Bug Fixes 🐛

• Fix Agent update planning logic by @Zoe14 in #1417
• Remove plural from named argument return_full_results in examples by @vladlen32230 in #1434
• Fix and refactor final answer checks by @aymeric-roucher in #1448
• Fix logging of Docker build logs by @tobiasofsn in #1421
• Add a mention of additional_args in the manager agent's prompt by @aymeric-roucher in #1441
• Fix missing mentions of additional_args in manager agent prompts by @albertvillanova in #1459
• Fix: __new__ method only accepts class object, remove **args and **kw… by @abdulhakkeempa in #1462
• Do not wrap types in safer_func by @albertvillanova in #1475
• Match multiline final answers in remote executors by @albertvillanova in #1444
• Revert removal of the last message from memory_messages during planning by @Zoe14 in #1454

Documentation Improvements 📚

• Quickstart Enhancements:

  • Add simple tool usage in doc quickstart by @aymeric-roucher in #1470
  • Add open-colab link to Quickstart docs by @albertvillanova in #1472

• Reference Docs:

  • Add AgentMemory to Reference docs by @albertvillanova in #1473
  • Add docstrings to GradioUI by @albertvillanova in #1451

• Fixes & Formatting:

  • Correct broken link in documentation homepage by @johntony366 in #1468
  • Rename docs files from .mdx to .md by @albertvillanova in #1471

Maintenance 🏗️

• Bump dev version: v1.19.0.dev0 by @albertvillanova in #1427

New Contributors

• @Zoe14 made their first contribution in #1417
• @vladlen32230 made their first contribution in #1434
• @JakeBx made their first contribution in #1420
• @abdulhakkeempa made their first contribution in #1462
• @johntony366 made their first contribution in #1468

Full Changelog: v1.18.0...v1.19.0

v1.18.0

New Features ✨

• Multiple Parallel Tool Calls: ToolCallingAgent can now handle multiple tool calls in parallel, significantly enhancing performance for complex tasks.
  • Support multiple tool calls in parallel in ToolCallingAgent by @albertvillanova in #1412
• Streaming Output for ToolCallingAgent: ToolCallingAgent now supports streaming outputs, improving responsiveness and user experience during multi-step tool interactions
  • Streaming outputs for ToolCallingAgent 🚀 by @aymeric-roucher in #1409
• API Web Search Tool: Introduced a new ApiWebSearchTool, enabling structured web search capabilities via API.
  • Create API web search tool by @albertvillanova in #1400

Enhancements 🛠️

• Update prompts to avoid confusion: search -> web_search and wiki -> wiki_search by @SrzStephen in #1403
• Make Agent.system_prompt read only by @albertvillanova in #1399
• Support configurable tool_choice in prepare_completion_kwargs by @albertvillanova in #1392
• Support passing additional params to MLXModel load and tokenizer.apply_chat_template by @albertvillanova in #1406
• Support custom headers/params for ApiWebSearchTool by @albertvillanova in #1411

Bug Fixes 🐛

• Fix: Support custom inputs execution for custom Final Answer Tool by @Lrakotoson in #1383
• Fix @tool decorator for remote Python executor by @tobiasofsn in #1334
• Always pass add_generation_prompt=True to apply_chat_template by @albertvillanova in #1416

Documentation Improvements 📚

• Make docs neutral about agent types by @julien-c in #1376
• Add MCP tools doc section by @aymeric-roucher in #1394
• Update ToolCollection.from_mcp docstring examples by @neonwatty in #1398
• Document better final_answer_checks by @albertvillanova in #1407
• Fix typo in secure code execution documentation by @chahn in #1414
• Add chat server demo to examples by @albertvillanova in #1415
• Fix VLM model in web_browser example docs by @albertvillanova in #1424
• Fix ValueError on Guided Tour docs by @albertvillanova in #1425
• Explain agent types better in Guided tour docs by @albertvillanova in #1426
• Add Quickstart page to docs by @albertvillanova in #1413

Maintenance 🏗️

• Bump dev version: v1.18.0.dev0 by @albertvillanova in #1390

New Contributors

• @Lrakotoson made their first contribution in #1383
• @neonwatty made their first contribution in #1398
• @SrzStephen made their first contribution in #1403
• @chahn made their first contribution in #1414

Full Changelog: v1.17.0...v1.18.0

v1.17.0

New Features ✨

• Structured Generation in CodeAgent: Add optional support for structured outputs in CodeAgent, enabling more reliable and consistent generation patterns
  • Adding optional structured generation to CodeAgent by @akseljoonas in #1346
• Support for Streamable HTTP MCP Servers: Expand compatibility with new server types to support streamable HTTP MCP implementations
  • Support Streamable HTTP MCP servers by @albertvillanova in #1384
• Run Results from Agent.run(): The Agent.run() method can now return a RunResult object, providing richer metadata on agent execution
  • Agent.run() can return RunResult object by @aymeric-roucher in #1337

Security Enhancements 🔒

• Safer LocalPythonExecutor: Prevent indirect submodule access via attribute resolution, mitigating potential security risks in user-defined code
  • Prevent submodules through indirect attribute access in LocalPythonExecutor by @albertvillanova in #1375

Enhancements 🛠️

• LocalPythonExecutor Enhancements:
  • Support annotated assignments within class definition in LocalPythonExecutor by @albertvillanova in #1355
  • Fix evaluate_class_def for Assign Attribute target in LocalPythonExecutor by @albertvillanova in #1357
  • Support 'pass' statement in class definition in LocalPythonExecutor by @albertvillanova in #1358
  • Refactor tests of LocalPythonExecutor by @albertvillanova in #1356
• Improve LaTeX rendering in GradioUI with extended delimiter support by @albertvillanova in #1387

Bug Fixes 🐛

• Streaming fixes
  • Fix live streaming when generating planning steps by @FlorianVal in #1348
  • Stop streaming if LiteLLM provide a finish_reason by @FlorianVal in #1350
  • add api_base and api_keys to preparation of kwargs for generate stream by @FlorianVal in #1344
• Fix WebSearchTool validation error by @albertvillanova in #1367
• Fix smolagents benchmark by @aymeric-roucher in #1377

Documentation Improvements 📚

• Add example docs about using OpenRouter models by @albertvillanova in #1364
• Fix Llama model name in docs example by @SaiDunoyer in #1379
• fix typo in docstring in mcp_client.py by @grll in #1380
• Document use_structured_outputs_internally with version added by @albertvillanova in #1388
• Fix rendering of version added in docs by @albertvillanova in #1389
• Fix broken link in agentic RAG examples page by @vksx in #1363

Maintenance 🏗️

• Bump dev version: v1.17.0.dev0 by @albertvillanova in #1336
• Remove deprecated from_hf_api methods by @albertvillanova in #1351

New Contributors

• @FlorianVal made their first contribution in #1344
• @vksx made their first contribution in #1363
• @akseljoonas made their first contribution in #1346
• @SaiDunoyer made their first contribution in #1379

Full Changelog: v1.16.1...v1.17.0

v1.16.1

Bug Fixes 🐛

• Revert "Add timeout mechanism for code execution" by @albertvillanova in #1340

Full Changelog: v1.16.0...v1.16.1

v1.16.0

New features 🚀

• Tools
  • Add Bing search engine to WebSearchTool by @albertvillanova in #1313
• Local Python executor improvements
  • Support passing custom functions to local Python executor by @albertvillanova in #1302
  • Support executor_kwargs in LocalPythonExecutor for initialization customization by @albertvillanova in #1314
  • Add timeout mechanism for code execution by @albertvillanova in #1318
• CLI improvements
• Enable local web agents via api_base and api_key by @zbloss in #1333

Important changes ⚠️

• Default provider for InferenceClientModel is now "auto"
  The default value of the provider argument in InferenceClientModel has changed from "hf-inference" to "auto".
  With "auto", the client will automatically select the first available provider for the model, following the user's configured priority at https://hf.co/settings/inference-providers
  • Pin minimum version of huggingface-hub to use InferenceClient auto provider by @julien-c in #1322

Bug Fixes 🐛

• Streaming GradioUI fixes
  • Fix thought yield in GradioUI for streaming and non-streaming by @albertvillanova in #1311
  • Fix duplicate error message in GradioUI, test, and refactor by @albertvillanova in #1312
  • Fix duplicate plan display in GradioUI when streaming by @albertvillanova in #1317

Documentation 📚

• Improve docs about Agentic RAG by @albertvillanova in #1319
• Update Readme video with streaming example by @aymeric-roucher in #1306

Misc 🛠️

• Bump dev version: v1.16.0.dev0 by @albertvillanova in #1304
• Update minimum supported Gradio version to 5.14.0 to fix Sidebar-related AttributeError by @albertvillanova in #1309
• Fix deprecation of HfApiModel by @albertvillanova in #1315
• Fix CI DuckDuckGoSearchTool tests raising 202 Ratelimit by @albertvillanova in #1324
• Remove utils dir from Makefile check_dirs by @tobiasofsn in #1335

New Contributors

• @tobiasofsn made their first contribution in #1335
• @zbloss made their first contribution in #1333

Full Changelog: v1.15.0...v1.16.0