v1.4.0

What's Changed

• fix(telemetry): added mcp tracing context propagation by @poshinchen in #569
• Change max_tokens type to int to match Anthropic API by @vinc3m1 in #588
• feat: Add additional intructions for contributors to find issues that are ready to be worked on by @mehtarac in #595
• feat(a2a): configurable request handler by @jer96 in #601
• chore(a2a): update host per AppSec recommendation by @jer96 in #619
• fix(event_loop): ensure tool_use content blocks are valid after max_tokens to prevent unrecoverable state by @dbschmigelski in #607
• fix(structured_output): do not modify conversation_history when prompt is passed by @dbschmigelski in #628

New Contributors

• @vinc3m1 made their first contribution in #588

Full Changelog: v1.3.0...v1.4.0

v1.3.0

What's Changed

• fix(deps): pin a2a-sdk>=0.2.16 to resolve #572 by @minorun365 in #581
• chore: pin a2a to a minor version while it is still in beta by @dbschmigelski in #586
• fix: sessions code fence, a2a tests & lint by @theagenticguy in #591
• fix(event_loop): raise dedicated exception when encountering max toke… by @dbschmigelski in #576

New Contributors

• @minorun365 made their first contribution in #581

Full Changelog: v1.2.0...v1.3.0

v1.2.0

What's Changed

• Support for Amazon SageMaker AI endpoints as Model Provider by @dgallitelli in #176
• fix: Remove leftover print statement from sagemaker model provider by @mehtarac in #553
• [Feat] Update structured output error message by @Unshure in #563
• feat(mcp): retain structured content in the AgentTool response by @dbschmigelski in #528
• feat(mcp): Add list_prompts, get_prompt methods by @Ketansuhaas in #160

New Contributors

• @dgallitelli made their first contribution in #176
• @mehtarac made their first contribution in #553
• @Ketansuhaas made their first contribution in #160

Full Changelog: v1.1.0...v1.2.0

v1.1.0

What's Changed

• Update to use dedicated github logo by @zastrowm in #505
• deps(a2a): address interface changes and bump min version by @jer96 in #515
• ci: expose STRANDS_TEST_API_KEYS_SECRET_NAME to integration tests by @dbschmigelski in #513
• Don't re-run workflows on un/approvals by @zastrowm in #516
• Doc fixes: suppressing some typos in various texts by @didier-durand in #487
• docs: add hot reloading documentation for load_tools_from_directory by @cagataycali in #517
• ci: enable integ tests for anthropic, cohere, mistral, openai, writer by @dbschmigelski in #510
• Automatically flatten nested tool collections by @zastrowm in #508
• feat(a2a): support mounts for containerized deployments by @jer96 in #524
• fix: include agent trace into tool for agent as tools by @poshinchen in #526

Full Changelog: v1.0.1...v1.1.0

v1.0.1

What's Changed

• Switch readme to use light logo for better display in github dark mode by @zastrowm in #475
• build(pyproject): update development status classifier by @awsarron in #480
• fix: enable parallel execution in graph workflow by @ahmetatalay in #485
• docs: Update README.md with Writer by @samjulien in #474
• fix(agent): prevent JSON serialization errors with non-serializable direct tool parameters by @cagataycali in #498
• fix(telemetry): group traces when using agent as tool in an agent by @poshinchen in #493

New Contributors

• @ahmetatalay made their first contribution in #485
• @samjulien made their first contribution in #474

Full Changelog: v1.0.0...v1.0.1

v1.0.0

What's Changed

• refactor(a2a): configurable host and port and remove excessive logging by @jer96 in #423
• models - bedrock - remove signaling by @pgrayy in #429
• feat(multiagent): Graph - support multi-modal inputs by @awsarron in #430
• deps(a2a): upper bound a2a sdk dep by @jer96 in #432
• models - ollama - init async client per request by @pgrayy in #433
• models - mistral - init client on every request by @pgrayy in #434
• models - ollama - clean up in tests by @pgrayy in #435
• feat!: add pagination to mcp_client list_tools_sync by @dbschmigelski in #436
• Session persistence by @Unshure in #302
• chore: update span names by @poshinchen in #440
• models - openai - null usage by @pgrayy in #442
• feat: redact content from a message in a session by @Unshure in #446
• refactor(a2a): upper bound deps + remove from multiagent submodule by @jer96 in #447
• Expand additional $refs for structured_output by @zastrowm in #439
• fix: session manager tracks all agent last message by @Unshure in #455
• feat: added swarm and graph spans by @poshinchen in #451
• docstrings - fix formatting by @pgrayy in #456
• refactor: add kwargs to multiagent interfaces by @mkmeral in #454
• fix: Fix session manager agent init by @Unshure in #458
• feat: Store conversation manager in session by @Unshure in #441
• feat(multiagent): introduce Swarm multi-agent orchestrator by @awsarron in #416
• multiagent - use invoke_async instead of stream_async by @pgrayy in #463
• feat: add Swarm tracing by @poshinchen in #461
• feat(telemetry): Expose OpenTelemetry exporter init arguments in API by @mrtj in #365
• docs: correct naming in registry.py by @akshseh in #425
• fix: Plumb system_prompt through to structured_output by @zastrowm in #466
• Update default model to be Claude 4 Sonnet by @zastrowm in #467
• feat: Add kwargs to session interfaces for future extensibility by @Unshure in #464
• fix: Fix various docstring issues by @Unshure in #469
• fix(multiagent): raise ValueError for unsupported Graph and Swarm agent features by @awsarron in #472
• refactor(multiagent): Swarm - Remove unnecessary complete_swarm_task tool by @awsarron in #473
• chore: remove preview from README.md by @yonib05 in #459

New Contributors

• @mrtj made their first contribution in #365
• @akshseh made their first contribution in #425

Full Changelog: v0.3.0...v1.0.0

v0.3.0

Breaking Changes

• Model Providers:

  • OpenAI and LiteLLM model providers no longer accept base64 encoded images. These providers will now handle base64 encoding of image bytes automatically. See issue #252.
  • Model providers now implement asynchronous functions. Custom model providers must be updated accordingly. See PR#306.
  • The API interface for model providers has been simplified to require only a single stream method implementation. See PR#400.
  • Model provider API interfaces have been relocated to the strands.models subpackage for improved organization. See PR#409.

• Other breaking changes:

  • All classes or protocols requiring subclassing or implementation now include an additional kwargs argument for forward compatibility. See PR#413.
  • The max_parallel_tools parameter has been removed from the Agent class. Tools are now automatically invoked on thread-pools. See PR#391.
  • The load_tools_from_directory parameter of the Agent class now defaults to False meaning. Tools are no longer automatically loaded from the tools directory. To restore the old behavior, explicitly set this parameter to True . See PR#419.
  • The event_loop_cycle function is no longer available as a top-level import from strands, reflecting its internal nature. Direct event loop invocation is not recommended. See PR#415.
  • Automatic removal of dangling tool messages during Agent lifecycle has been discontinued to support future features. See PR#418.
  • AgentTool now implements a stream API instead of invoke and requires asynchronous implementation. See PR#345.

What's Changed

• feat: Implement the core system of typed hooks & callbacks by @zastrowm in #304
• iterative tool handler process by @pgrayy in #340
• remove thread pool wrapper by @pgrayy in #339
• chore: updated scope name, enable setting up meter by @poshinchen in #331
• async model stream interface by @pgrayy in #306
• chore: allow custom agent name by @poshinchen in #347
• Extract hook based tests to a separate file by @zastrowm in #349
• tools - parallel execution - sleep by @pgrayy in #355
• Refactor event loop to use Agent object rather than individual parameters by @zastrowm in #359
• models - openai - async client by @pgrayy in #353
• models - openai - do not accept b64 images by @pgrayy in #368
• iterative tools by @pgrayy in #345
• a2a streaming by @jer96 in #366
• docs(multiagent): Update A2AServer docstrings by @signoredems in #377
• refactor(a2a): move a2a test module by @jer96 in #379
• models - mistral - async by @pgrayy in #375
• models - ollama - async by @pgrayy in #373
• models - anthropic - async by @pgrayy in #371
• feat: Add hooks for before/after tool calls + allow hooks to update values by @zastrowm in #352
• agent tool - remove invoke by @pgrayy in #369
• fix: handle multiple tool calls in Mistral streaming responses by @siddhantwaghjale in #384
• fix: add-threading-instrumentation by @poshinchen in #394
• Add cohere client by @billytrend-cohere in #236
• deps(a2a): upgrade a2a with db support by @jer96 in #395
• Writer model provider by @yanomaly in #228
• Update integ tests to isolate provider-based tests by @zastrowm in #396
• Remove agent.tool_config and update usages to use tool_specs by @zastrowm in #388
• multi modal input by @pgrayy in #367
• Fix: Update mistral tests to avoid shared agents by @zastrowm in #398
• async tools support by @pgrayy in #391
• Add basis for conformance-based tests by @zastrowm in #403
• fix: Allow tool names that start with numbers by @zastrowm in #407
• Add hooks for when new messages are appended to the agent's messages by @zastrowm in #385
• Add Model Invocation Hooks by @zastrowm in #387
• structured output - multi-modal input by @pgrayy in #405
• feat(async): mcp async call tool by @jer96 in #406
• [REFACTOR] Unify Model Interface Around Single Entry Point (model.stream) by @mkmeral in #400
• Rename StartRequestEvent & EndRequestEvent events by @zastrowm in #408
• models - bedrock - threading by @pgrayy in #411
• Mark hooks as non-experimental by @zastrowm in #410
• models - litellm - async by @pgrayy in #414
• models - move abstract class by @pgrayy in #409
• feat(multi-agent): introduce Graph multi-agent orchestrator by @awsarron in #336
• refactor: Remove event_loop_cycle from top level import by @Unshure in #415
• refactor: Remove message processor by @Unshure in #417
• refactor: Update interfaces to include kwargs to enable backwards compatibility by @mkmeral in #413
• refactor: Remove _remove_dangling_messages from SlidingWindowConversationManager by @Unshure in #418
• chore!: set Agent property load_tools_from_directory to default to False by @dbschmigelski in #419

New Contributors

• @signoredems made their first contribution in #377
• @billytrend-cohere made their first contribution in #236
• @yanomaly made their first contribution in #228
• @mkmeral made their first contribution in #400

Full Changelog: v0.2.1...v0.3.0

v0.2.1

What's Changed

This is a hotfix for v0.2.0

• tools - parallel execution - sleep by @pgrayy in #355

Full Changelog: v0.2.0...v0.2.1

v0.2.0

Minor Version Bump to v0.2.0

For this minor version bump, we are introducing a number of backwards-incompatible changes. Leading up to the 1.0.0 release of the SDK, we are incrementally releasing backwards-incompatible changes to align with our vision of the 1.0.0 interface. These changes are meant to fix deficiencies that were part of our initial implementation, or to make Strands easier to understand and use. Below are a list of these changes, and steps to migrate your code if they affect you:

• Use region from boto3 session when possible (#299)
  • This change updates the region behavior of the BedrockModelProvider to better align with the Boto3 default region behavior instead of us-west-2.
• refactor: remove kwargs spread after agent call (#289)
  • Previously, strands used certain parameters that were passed into the agent invocation call. For example, agent("Hello!", system_prompt="Some New Prompt") would previously override the system prompt defined during the agent’s initialization. Now, this behavior has been removed, so you can alter the agent’s system prompt by overriding the agents system_prompt attribute.
    • Before: agent("Hello!", system_prompt="Some New Prompt")
    • After: agent.system_prompt = "Some New Prompt"
• Remove FunctionTool (#325)
  • Previously, FunctionTool was deprecated in favor of DecoratedFunctionTool but it was kept in for backwards compatibility. This removes the FunctionTool class in favor of using the @tool decorator. Check out the strands agents tool documentation for instructions on building a tool.
• Implement native async iterator support (#83)
  • #295: We no longer yield any callback events carried out in tool invocations from stream_async.
  • #323: callback_handler is no longer passed to tools.
• refactor tracer (#286)
  • The OTEL setup process has changed. Before, Strands would try to automatically setup and OTEL exporter using the STRANDS_OTEL_ENABLE_CONSOLE_EXPORT environment variable. This has been removed, and replaced with a function call. See the docs here for more info: https://strandsagents.com/latest/user-guide/observability-evaluation/traces/#code-configuration

What's Changed

• iterative event loop by @pgrayy in #268
• Add additional exception information for common bedrock errors by @zastrowm in #290
• iterative structured output by @pgrayy in #291
• fix: correcting incorrect docstring in tracer.py - non-existing argument documented by @didier-durand in #293
• tools - do not remove $defs by @pgrayy in #294
• feat: Add reasoning content for openai model provider by @xiehust in #187
• chore: refactor tracer by @poshinchen in #286
• iterative agent by @pgrayy in #295
• feat(a2a): tools as skills by @jer96 in #287
• feat: Add Mistral model support to strands by @siddhantwaghjale in #284
• fix: Fix docs warnings by @Unshure in #303
• Use region from boto3 session when possible by @zastrowm in #299
• chore: update spanKind and attributes for tokens by @poshinchen in #296
• fix: Migrate Mistral structured_output to an iterator by @zastrowm in #305
• feat: add debug logging for model converse requests by @RingoIngo2 in #297
• refactor: remove kwargs spread after agent call by @Unshure in #289
• chore: allow custom tracer_provider and chain setup by @poshinchen in #316
• feat: Add reproduction test for #320 by @zastrowm in #322
• feat: Agent State by @Unshure in #292
• stop passing around callback handler by @pgrayy in #323
• refactor: Remove unused code by @Unshure in #326
• chore: updated semantic conventions on Generative AI spans by @poshinchen in #319
• refactor: Consolidate agent state unit tests by @Unshure in #334
• Remove FunctionTool as a breaking change by @zastrowm in #325
• executor - run tools - yield by @pgrayy in #328

New Contributors

• @siddhantwaghjale made their first contribution in #284
• @RingoIngo2 made their first contribution in #297

Full Changelog: v0.1.9...v0.2.0

v0.1.9

What's Changed

• feat: add meter by @poshinchen in #219
• models - openai - images - b64 validate by @pgrayy in #251
• chore: Inline event loop helper functions by @zastrowm in #222
• feat: add structured output support using Pydantic models by @theagenticguy in #60
• fix: Emit warning that default region behavior will be changing by @zastrowm in #254
• models - openai - b64encode method by @pgrayy in #260
• chore/update metrics by @poshinchen in #248
• iterative streaming by @pgrayy in #241
• Initial A2A server Integration by @jer96 in #218
• litellm - bug in v1.73.0 by @pgrayy in #270
• Update @tool to return an AgentTool that also acts as a function by @zastrowm in #258

New Contributors

• @theagenticguy made their first contribution in #60

Full Changelog: v0.1.8...v0.1.9