v1.7.0

Main changes

Added commands to directly run agents from CLI: smolagent and webagent ✨

• Make web browser agent example a CLI script by @merveenoyan in #416
• Add CLI for smolagents by @aymeric-roucher in #431
  • Now you can run smolagent "Your task!" from CLI to directly call a smolagent. Try it! webagent "Find me a cheap train from Paris to Torino before Thursday" or smolagent "Who's the current director of La French Tech?"

Gave a memory to agents! 🧠

• Untangling Logging by @clefourrier in #316
• Test memory by @albertvillanova in #428
  • This added a specific memory attribute to the agent!
  • We used this change to introduce nice functions like an agent.replay() function that replays the last run (based on the memories stored, so no further LLM calls)

Other big improvements 💪

• Improve code execution logs in case of error by showing print outputs by @aymeric-roucher in #446

  • This makes sure that code execution outputs are stored to memory even when an error was raised later on. It should slightly improve the performance of CodeAgents

• Update benchmark with Hub datasets by @albertvillanova in #412

  • This adds storing of benchmarking outputs directly to the Hub. Moving forward, we'll make benchmarking even easier!

Fixes and small improvements

• Bump dev version: v1.7.0.dev0 by @albertvillanova in #390
• Fix return type of generator function in Agent._run() by @abidlabs in #409
• Update minimum Gradio version by @abidlabs in #408
• Support third-party Inference providers in HfApiModel by @julien-c in #422
• Minor fix for agent_from_any_llm.py example by @RolandJAAI in #423
• Fix vlm_web_browser.py example by @abidlabs in #410
• Fix regressions, refactor logging and improve replay function by @aymeric-roucher in #419
• Fix installation instruction for GradioUI in error message by @jank in #418
• Fix import from deleted logger module by @albertvillanova in #433
• Add web search tool and make prompt positional arg by @aymeric-roucher in #435
• Remove explicit model arg in HfApiModel - which wouldn't run without Pro by @sanjeed5 in #373
• Fix MultiStepAgent.planning_step message content by @albertvillanova in #437
• Fix and test MemoryStep by @albertvillanova in #432
• Fix get safe module by @antoinejeannot in #405
• Remove duplicate final_memory_step creation by @albertvillanova in #444
• Remove unused ChatMessageToolCall raw argument by @albertvillanova in #445
• Fix TransformersModel docstring by @August-murr in #447
• Test get_clean_message_list by @albertvillanova in #448
• Add web browser example to doc by @aymeric-roucher in #439
• Fix minor issue in e2b by @keetrap in #436

New Contributors

• @abidlabs made their first contribution in #409
• @julien-c made their first contribution in #422
• @August-murr made their first contribution in #447

Full Changelog: v1.6.0...v1.7.0

v1.6.0

What's Changed

• Bump version to 1.6.0.dev by @albertvillanova in #348
• Use transformers as optional dependency by @albertvillanova in #360
• Add pip install datasets to RAG example by @sanjeed5 in #371
• Update issue templates by @aymeric-roucher in #377
• TransformersModel auto-detects VLMs by @aymeric-roucher in #378
• Separate tree for Model docs is added by @touseefahmed96 in #382
• add kwargs to gradio launch by @tocab in #367
• Minor Fix in the example docs by @touseefahmed96 in #383
• Fix Extra planning by @bvantuan in #362
• Gradio chatbot: step duration, number, token count, support nested thoughts by @yvrjsharma in #384
  • Gradio Chatbot looks much nicer than before! See below:

• Add Chinese translations for text-to-sql example by @FanwangM in #366
• Add Chinese translations for RAG by @FanwangM in #363
• Qwen-2VL -> Qwen2-VL model name typo fixed by @sergiopaniego in #349
• chore: Fix Typo by Calling is_torch_available() in AgentAudio Class by @hironow in #379
• Additional parameters for openai client by @touseefahmed96 in #385
• DuckDuckGoSearchTool: add ddgs_kwargs parameter to constructor by @onukura in #372
• Fix final_answer issue in e2b_executor by @keetrap in #319
• Docs: Text to sql example - minor naming / dedupe by @GrahamTheCoder in #376
• Test import without extras by @albertvillanova in #370
• Add tool saving test by @aymeric-roucher in #389
• Fix blocking of os in authorized imports by @aymeric-roucher in #386
• Add test for DDGS Tool kwargs by @aymeric-roucher in #387

New Contributors

• @tocab made their first contribution in #367
• @bvantuan made their first contribution in #362
• @sergiopaniego made their first contribution in #349
• @hironow made their first contribution in #379
• @onukura made their first contribution in #372
• @GrahamTheCoder made their first contribution in #376

Full Changelog: v1.5.1...v1.6.0

v1.5.1

Bug Fixes 🐛

• Use transformers as optional dependency by @albertvillanova in #360

Full Changelog: v1.5.0...v1.5.1

v1.5.0

What's Changed

New Features 🚀

• Add VLM support by @merveenoyan in #220

Fixes 🐛

• refactor: update model type to ChatMessage in agent classes by @jank in #263
• Fixes AttributeError: 'CodeAgent' object has no attribute 'logger' #258 by @joaopauloschuler in #259
• Improve tool call argument parsing by @aymeric-roucher in #267
• Remove unused and undocumented test_mode parameter by @Wauplin in #273
• Fix Bug in from_langchain in tools.py by @RolandJAAI in #268
• Python interpreter: improve suggestions for possible mappings by @aymeric-roucher in #266
• Improve python executor's error logging by @aymeric-roucher in #275
• The use of BASE_PYTHON_TOOLS enables the use of compile, exec and eval by @tandiapa in #274
• Try first dunder method in evaluate_augassign by @albertvillanova in #285
• Add cool GIF of agent run inspection using Phoenix by @aymeric-roucher in #277
• Remove pickle5 package from E2BExecutor by @albertvillanova in #295
• Fuse stream and direct run calls by @clefourrier in #296
• minor fix for console in AgentLogger by @nbroad1881 in #303
• Add Azure OpenAI support by @vladiliescu in #282
• Fix arg passing to AgentExecutionError by @albertvillanova in #309
• refactor(models): restructure model parameter handling by @kingdomad in #227
• Local interpreter security: prevent builtins functions from being used if they have not been added as tools by @tandiapa in #299
• Support any and none tool types by @aymeric-roucher in #280
• Minor fix: adding a 60 seconds timeout to the visit webpage tool by @Killian-pit in #308
• Fix: source code inspection in interactive shells by @antoinejeannot in #281
• Refactor evaluate_augassign and test all operators by @albertvillanova in #313
• Unset temperature in models by @aymeric-roucher in #315
• Fix tool calls with LiteLLM and tool optional types by @aymeric-roucher in #318
• Multiple tool example by @touseefahmed96 in #293
• RAG on your huggingface_doc data using chromadb and groq api by @touseefahmed96 in #235
• Improve static tools initialization safety by @kingdomad in #324
• Append unused error message to memory by @albertvillanova in #325

Documentation and examples 📚

• Fix code examples with additional_args and num_ctx examples also in zh docs by @RolandJAAI in #260
• Update building_good_agents.md by @derekalia in #283
• Update guided_tour.md by @sanjeed5 in #287
• Added Hindi docs for smolagents by @keetrap in #279
• Update README and documentation to clarify Hub integrations with Gradio Spaces by @davidberenstein1957 in #306
• Update README.md fix quick demo code import bug by @Deng-Xian-Sheng in #327
• Add args to MultiStepAgent docstring by @albertvillanova in #332
• Fix MultiStepAgent docstring by @albertvillanova in #336
• Fix docstrings of models by @albertvillanova in #344
• docstring args for ToolCallingAgent, CodeAgent and ManagedAgent by @touseefahmed96 in #335
• Corrected tool examples in ToolCallingAgent system prompts by @RolandJAAI in #341
• Improve inference choice examples by @aymeric-roucher in #311

Maintenance 🏗️

• Add linter rules + apply make style by @Wauplin in #255
• Fix CI quality issue by @albertvillanova in #272
• Bump version to 1.5.0.dev by @aymeric-roucher in #256
• Make import time faster (optional deps + delay imports) by @Wauplin in #253
• Fix CI quality by @albertvillanova in #286
• Add huggingface-hub as required dependency by @albertvillanova in #288
• Make e2b optional dependency by @albertvillanova in #292
• Remove package json files by @albertvillanova in #314
• Update README instructions to run tests by @albertvillanova in #328
• Move torchvision to the torch extra by @nonsleepr in #297

New Contributors

• @tandiapa made their first contribution in #274
• @keetrap made their first contribution in #279
• @sanjeed5 made their first contribution in #287
• @davidberenstein1957 made their first contribution in #306
• @clefourrier made their first contribution in #296
• @nbroad1881 made their first contribution in #303
• @vladiliescu made their first contribution in #282
• @Killian-pit made their first contribution in #308
• @antoinejeannot made their first contribution in #281
• @touseefahmed96 made their first contribution in #293
• @Deng-Xian-Sheng made their first contribution in #327
• @nonsleepr made their first contribution in #297
• @merveenoyan made their first contribution in #220

Full Changelog: v1.4.1...v1.5.0

v1.4.1

What's Changed

New features 🚀

• Add support for MCP Servers tools as ToolCollection by @grll in #232
  • This enables using Model Context Protocol (MCP) servers, thus makes many, many tools compatible with smolagents: find a list on glama.ai or smithery.ai ✨
• Allow passing kwargs to all models by @aymeric-roucher in #222
• Add Chinese translations for multi-agent example by @FanwangM in #246

Fixes 🐛

• Fix subpackage import vulnerability by @aymeric-roucher in #238
• Fix typo in building_good_agents.md by @rug in #193
• Align data types in example benchmark by @albertvillanova in #205
• Rename the split of the benchmark dataset from train to test by @albertvillanova in #216
• Fix vanilla model answer in example benchmark by @albertvillanova in #219
• Fix additional args in stream_to_gradio by @aymeric-roucher in #221
• Return textboxes on Gradio file upload errors by @stackviolator in #214
• Docs Bug - fix tool example with additional args by @RolandJAAI in #228
• Dead Link to Duck Duck Go search tool by @matterattetatte in #233
• Fix missing python modules in CodeAgent system prompt by @edbeeching in #226

Adjustments 🔧

• Clean makefile, pyproject.toml and CI by @Wauplin in #229
• Add resizeable option to Gradio UI component for better usabilty by @jank in #234
• Add trust_remote_code arg to TransformersModel by @aymeric-roucher in #240
• Make openai dependency optional by @albertvillanova in #236
• Remove dependency on _is_package_available from transformers by @albertvillanova in #247
• Update README.md by @aymeric-roucher in #208
• Set ollama context length in example to 8192 to make it work by @RolandJAAI in #252
• chore: unset console width, utilize auto detect feature by @simpletrontdip in #196
• Max length of "print" outputs as a parameter of an agent by @IlyaGusev in #209
• Small fixes in docs by @MoritzLaurer in #245
• Fixes before release by @aymeric-roucher in #254

New Contributors

• @rug made their first contribution in #193
• @RolandJAAI made their first contribution in #228
• @matterattetatte made their first contribution in #233
• @jank made their first contribution in #234
• @edbeeching made their first contribution in #226
• @Wauplin made their first contribution in #229
• @FanwangM made their first contribution in #246
• @simpletrontdip made their first contribution in #196
• @MoritzLaurer made their first contribution in #245
• @grll made their first contribution in #232

Full Changelog: v1.3.0...v1.4.1

v1.3.0

What's Changed

New features 🚀

• We now support OpenTelemetry instrumentation! Find a detailed tutorial here 📊 by @aymeric-roucher in #200

• We now have a Chinese (mandarin) version of the doc for smolagents! by @Ayuilos in #156

• Add option to upload files to GradioUI by @stackviolator in #138

• feat: Add multi-GPU support for TransformersModel by @6643789wsx in #139

• More flexible verbosity level: now use verbosity_level=0/1/2 instead of verbose=True/False upon agent initialization by @aymeric-roucher in #150. This gives you more fine-grained control over verbosity!

• Agents can now import anything using additional_authorized_imports=['*']! Contributed by @joaopauloschuler in #129

Fixes and smol improvements 🔧

• Add top level docs link to README by @robtaylor in #121
• Added the right link for website button by @AngeloKiriakoulis in #159
• Tools from Spaces: Fix bug preventing use of long prompts. by @jkingston in #134
• Fix typo in dosctrings by @chloefeal in #155
• Log list of tool calls in ActionStep by @aymeric-roucher in #172
• Bug fixes on TransformersModel by @AngeloKiriakoulis in #165
• Fix wrong return value in agents.md doc by @sidtuladhar in #164
• Fix minor issues in docs by @duydl in #173
• Fix quality issues by @albertvillanova in #171
• Clean local python interpreter: propagate imports by @aymeric-roucher in #175
• Fix tool_calls parsing error in ToolCallingAgent when using OpenAIServerModel by @tanhuajie in #160
• Fix minor issues in building_good_agents docs by @albertvillanova in #170
• bugfix: Fix plan_update message display by @IlyaGusev in #179
• Make default tools more robust by @aymeric-roucher in #186
• Implemented support for ast.Pass in the interpeter. by @AngeloKiriakoulis in #189
• Support pandas' iloc indexer by @aymeric-roucher in #191
• The call.func parameter type of the local_python_executor.evaluate_call function might be ast.Subscript by @nvrxq in #194
• bug fix: fix string concatenation bug in GradioUI.log_user_message by @kingdomad in #199
• Fix import from ChatMessage in test_monitoring by @aymeric-roucher in #202
• Fix quality in gradio_ui by @albertvillanova in #206
• Pre-release fixes by @aymeric-roucher in #207
• Document OpenTelemetry usage by @aymeric-roucher in #204

New Contributors

• @robtaylor made their first contribution in #121
• @AngeloKiriakoulis made their first contribution in #159
• @jkingston made their first contribution in #134
• @chloefeal made their first contribution in #155
• @sidtuladhar made their first contribution in #164
• @Ayuilos made their first contribution in #156
• @duydl made their first contribution in #173
• @stackviolator made their first contribution in #138
• @albertvillanova made their first contribution in #171
• @tanhuajie made their first contribution in #160
• @IlyaGusev made their first contribution in #179
• @6643789wsx made their first contribution in #139
• @nvrxq made their first contribution in #194
• @kingdomad made their first contribution in #199

Full Changelog: v1.2.2...v1.3.0

v1.2.0

What's Changed

Big features:

• Add an OpenAIServerModel by @Zetaphor in #109
  • This unlocks using any OpenAI-format-compatible server like TGI or vLLM.
• Enable smolagent tools in Hugging Chat! 🚀 in #132
  • Demo incoming!
• Vastly simplify Model class ✨ in #146
  • Now Model subclasses have only one __call__ method: if arg tools_to_call_from is passed, this method will return a tool call instead of just a string.
• Halve import time by removing torch dependency in #147
  • Following up on #100 raised by @xianminx, removing dependency on torch allowed to halve the library loading time!
• Support OpenTelemetry in #136
  • This is still cooking, more stuff needs to be done, but we'll have observability!

Other important features

• Include missing pip install library in RAG example by @DarwinAnim8or in #92
• Update rag.md with missing verb by @derekalia in #95
• Small tweaks to Gradio demo by @freddyaboulton in #93
• Fixing spelling and grammar for query by @omarirfa in #98
• Fix Spelling Errors in Documentation by @ScientistIzaak in #99
• Change DuckDuckGoSearchTool output_type to string by @aymeric-roucher in #116
• Auto correct wrong assignments to final_answer by @aymeric-roucher in #123
• Adding "complex" to local_python_executor.py #78 by @joaopauloschuler in #79
• Remove dependency on LiteLLM by @aymeric-roucher in #126
• Support multiple code blobs by @aymeric-roucher in #128
• Fix several typos in docs. by @Symbolk in #140

Great thanks for everyone who participated!

v1.1.0

What's Changed

New features: 🥳

• !Breaking change: Agent initialization argument max_iterations changed to max_steps for consistency.
• Add a warning towards the user when imports are missing in CodeAgent 3f79bae.
• Add code of conduct and contributing guide by @LysandreJik in #50
• feat: Add max_results kwarg to DDGS tool by @Stillerman in #40
• Add support for additional keyword arguments in LiteLLMModel by @chakib-belgaid in #29
• Add device parameter for TransformerModel in models.py by @ScientistIzaak in #49

Fixes: 🔧

• Don't pass sanitize_inputs_outputs=True to managed agents by @grrowl in #85
• Include missing pip install library in RAG example by @DarwinAnim8or in #92
• Fix issue with forward signature validation in space tool wrapper by @lelayf in #75
• fix ToolCollection usage by @Brunwo in #53

Typos: 🪶

• Update building_good_agents.md by @SHUBH4M-KUMAR in #43
• Fix example usage in HfApiModel by @balikasg in #37
• Fixing minor spelling errors in building_good_agents.md by @ScientistIzaak in #33
• chore: update local_python_executor.py by @eltociear in #64
• Add missing end of sentence to building_good_agents by @elroy-bot in #68
• Fix typos by @oliveredget in #74
• Update README.md: add link to blog post by @gary149 in #22
• e2b details by @CakeCrusher in #72
• Fixed typos and grammatical errors in building_good_agents.md tutorial by @alxfgh in #89

Full Changelog: v1.0.0...v1.1.0

v1.0.0

Bump release number for 1.0.0

v0.1.0: First ever release!

Nits in readme