agents 2.0.0

Dart port of the Microsoft Agents AI framework. Build, compose, and orchestrate AI agents backed by any ChatClient — with sessions, streaming, tool use, compaction, evaluation, and middleware pipelines.

Changelog

2.0.0

• Breaking: streamAsync and resumeStreamAsync now return a live run. Both used to drive the workflow to quiescence before returning, so the returned StreamingRun was already finished and its outgoingEvents was complete. They now return as soon as the run is open and drive it in the background. Observe progress with StreamingRun.watchStreamAsync(); reading outgoingEvents immediately after the await yields only the events produced so far. External responses sent via sendResponseAsync (and messages via trySendMessageAsync) resume the run.
  • In ExecutionMode.offThread (the default) every event is published the moment it is created, including outputs yielded part-way through an executor invocation — streamed agent updates now surface live. In ExecutionMode.lockstep events are still batched and published together after each superstep.
  • watchStreamAsync replays the events already recorded before switching to the live tail, so a watcher that subscribes late — even after the run has ended — still observes the full sequence.
  • Added StreamingRun.isCompleted.
• Breaking: external responses route by request, not by payload type. addExternalResponse previously completed the request and then handed the payload to the first executor whose input type accepted it. It now resolves the pending ExternalRequest by requestId (searching joined sub-workflow runners too) and delivers the response to the executor that issued it. It throws a StateError — leaving pending requests untouched — when the request id is unknown, already serviced, or the issuing executor does not accept the payload type. Workflows that relied on the old loose matching must respond with the ExternalRequest surfaced by the run's RequestInfoEvent.
  • ExternalRequest gained sourceExecutorId, recording the issuing executor.
• WorkflowContext.sendRequest now returns ExternalResponse.pending, an explicit placeholder, instead of forcing null into TResponse. Reading response on a placeholder throws a StateError naming the request and port rather than failing with an opaque cast error; the real response arrives as a message in a later superstep.
• Added WireMarshaller.valueConverters, a registry of WireValueConverters keyed by payload type id. Checkpointing serializes pending message payloads to JSON text, which fails for payloads outside the JSON model (chat messages, for example); hosts can now register a converter to serialize and revive those types.
• AsyncRunHandle serializes concurrent drive passes, so a caller that delivers a message or response while a drive is in flight waits for the queued work to be processed. A failure in the initial background drive is now reported as a WorkflowErrorEvent and ends the run instead of leaving stream observers waiting forever.
• Breaking: requires extensions: ^0.6.0, which changes AI function invocation. Reaching maximumIterationsPerRequest now issues one more provider call with function declarations withheld so the model produces a real answer, instead of returning unanswered tool calls; exceeding maximumConsecutiveErrorsPerRequest now throws (a single failure rethrown as-is, several combined into an AggregateException) instead of returning a partial response; and the limit comparison changed from >= to >, so a limit of 0 surfaces the first tool failure immediately. Also inherits AIContent.annotations and the completed OpenTelemetry decorator set.
• Breaking: requires anthropic_sdk_dart: ^6.0.0. Its types appear in this package's public API (AnthropicChatClient.client and the AnthropicClient builder extensions), so dependents pinning the SDK directly must move to 6.x as well.
• Verified against the current releases of the unchanged constraints, notably mcp_dart 2.4.0.

1.6.0

• Add live shell output reporting: ShellExecutor.outputEvents broadcasts typed ShellOutputChunk values while commands run, local and Docker options accept an onOutput callback, and persistent shell sessions now capture stderr instead of discarding it.

1.5.0

• Add AnthropicChatClient under anthropic/: a ChatClient backed by Anthropic's Messages API, with client builder extensions and shared defaults.
• Add GeminiChatClient and GeminiClient under gemini/: a ChatClient backed by the Gemini API, with builder extensions and defaults. Handles thoughtSignature round-tripping on function calls (including the documented skip-validation placeholder for replayed calls), strips unsupported additionalProperties from tool and response schemas, and enables server-side tool invocations when mixing Gemini built-in tools with function tools.
• Add MCP integration under mcp/: AgentMcpSkillsSource discovers Agent Skills exposed over MCP (AgentMcpSkill, McpSkillIndex, AgentMcpSkillResource, plus options), McpClientTaskExtensions exposes MCP tools as AI functions (McpClientAIFunction, TaskAwareMcpClientAIFunction, McpTaskOptions), and the skills provider builder gains MCP registration extensions.
• Add sequential and concurrent orchestrations to the workflows engine: SequentialWorkflowBuilder and ConcurrentWorkflowBuilder on the shared OrchestrationBuilderBase, with OutputTag (+ JSON converter), WorkflowOutputEvent extensions, and checkpoint support for output executors in WorkflowInfo.
• Add session-store composition under hosting/: DelegatingAgentSessionStore, IsolationKeyScopedAgentSessionStore (+ options), and SessionIsolationKeyProvider for partitioning stored sessions by user, tenant, or composite keys.
• Expand OpenAI hosting: HostedAgentResponseExecutor routes Responses API requests to hosted agents by agent.name or metadata["entity_id"], with OpenAIResponseRequestInfo / OpenAIChatCompletionRequestInfo request descriptors, per-API map options, AgentReference, and standardized response error codes.
• Add OpenAI conversion helpers under ai/open_ai/: extensions mapping AgentResponse and ChatClientAgent results onto OpenAI wire formats.
• Add auto-approval rules to the tool-approval middleware: ToolApprovalAgentOptions with ordered autoApprovalRules (evaluated after standing rules, before prompting the user) and an approve-all rule.
• Expand the file access and file memory providers: new replace and replace_lines editing tools backed by the shared FileEditor / FileLineEdit helpers, FileStoreEntry metadata, configurable tool names, and read-only / full auto-approval rule presets.
• Add chat-client decorators: MessageInjectingChatClient lets external code (such as tool delegates) enqueue messages into the function invocation loop, and NonApprovalRequiredFunctionBypassingChatClient strips approval requests for tools that do not require approval, re-injecting them pre-approved on the next request.
• Add BackgroundTaskCompletionLoopEvaluator (+ options): keeps a LoopAgent iterating until tracked background tasks complete, with a templated feedback message listing the still-running tasks.
• Add skills improvements: a CachingAgentSkillsSource decorator (+ options), AgentSkillsSourceContext, and AgentFileSkillFilterContext for filtering file-based skills.
• Add evaluation types: GeneratedEvaluatorRef (versioned references to generated evaluators) and RubricScore (typed per-dimension score breakdown for rubric evaluators).
• Export InvokedContext and InvokingContext as standalone libraries (previously hidden from the public API) and publicly export ChatMessageJsonConverter.
• Fix AIContextProvider.getService to resolve requests for concrete provider types (matching the runtimeType idiom used by AgentSession and DelegatingAIAgent); concrete-type lookups previously returned null, breaking provider resolution in BackgroundTaskCompletionLoopEvaluator and TodoCompletionLoopEvaluator.
• Add anthropic_sdk_dart: ^5.0.0, archive: ^4.0.9, http: ^1.6.0, and mcp_dart: ^2.2.1 dependencies; bump extensions to ^0.5.0.

1.4.0

• Add loop agents under harness/loop/: LoopAgent and LoopAgentOptions run an inner agent repeatedly until an evaluator signals completion, with pluggable LoopEvaluator strategies — AIJudgeLoopEvaluator, CompletionMarkerLoopEvaluator, TodoCompletionLoopEvaluator, and DelegateLoopEvaluator — plus LoopContext, LoopEvaluation, and JudgeVerdict support types.
• Add Magentic multi-agent orchestration to the workflows engine: MagenticWorkflowBuilder, MagenticOrchestrator, the plan-review request/response messages, and the progress ledger (ports the upstream Magentic manager/orchestrator pattern).
• Add OpenAI-compatible hosting under hosting/open_ai/: shelf-based routers and handlers for the Chat Completions, Conversations, and Responses APIs, backed by in-memory storage, exposed through open_ai_hosting_service_collection_extensions.
• Add shelf: ^1.4.0 and shelf_router: ^1.1.0 dependencies for the OpenAI hosting routers.
• Fix fan-in edges losing buffered messages across checkpoint/resume: pending fan-in contributions are now captured in Checkpoint.fanInState and restored on resume (both the in-proc and legacy execution engines). Old checkpoint JSON without the field remains loadable.
• Fix fan-in edges dropping all but the last message from a source that sent more than once before the edge released; all buffered messages are now delivered, ordered by source then arrival (matches upstream FanInEdgeState semantics).
• Fix streamed responses losing responseId, messageId, createdAt, usage, modelId, rawRepresentation, and additionalProperties when coalesced by ChatClientAgent and PerServiceCallChatHistoryPersistingChatClient; all call sites now share one toChatResponse() extension (ports C# ToChatResponse).
• Fix A2AAgentSession.serialize() dropping the session stateBag; it now round-trips, and legacy payloads without it remain loadable.
• Remove the unused StatefulEdgeRunner interface (breaking; it had no implementors — fan-in state is checkpointed via Checkpoint.fanInState).
• Rename agent_response_t_.dart → agent_response_of.dart and provider_session_state_t_state_.dart → provider_session_state.dart (library paths only; type names unchanged).
• Simplify shell executor timeout handling with a shared waitForProcessExit helper; removes an unmanaged kill timer and ensures the force-kill is awaited before draining output.

1.2.0

• Add A2A (Agent-to-Agent) protocol support:
  • Client-side A2AAgent, A2AAgentOptions, A2AAgentSession, and A2AContinuationToken for consuming remote agents over the A2A protocol.
  • a2a_client_extensions and a2a_agent_card_extensions helpers.
  • Server-side hosting bridge: A2AAgentHandler, A2ARunDecisionContext, A2AServerRegistrationOptions, agentRunMode, the a2a_server_service_collection_extensions registration helpers, and a MessageConverter.
• Add a2a: ^4.2.0 dependency.

1.1.0

• Previous release.