TemporalCommunity.Extensions.Agents 0.12.1

TemporalAgents

Temporal .NET SDK integrations for building durable AI applications. Two packages, two levels of abstraction:

Both packages make their supported session and activity boundaries durable. Conversation history and LLM calls are persisted in Temporal history; registered durable tool invocations are separate activities and replay safely after crashes or restarts.

Overview

TemporalCommunity.Extensions.AI

A lightweight integration for Microsoft.Extensions.AI (MEAI). It provides direct IChatClient middleware for custom workflows and DurableChatSessionClient for managed multi-turn sessions. Managed sessions own their model/tool loop and use AddDurableTools for durable function dispatch. No Agent Framework dependency is required.

Start here if: you are already using MEAI's IChatClient directly and want Temporal durability without adopting the full Agent Framework.

dotnet add package TemporalCommunity.Extensions.AI

Full documentation →

TemporalCommunity.Extensions.Agents

A Temporal integration for Microsoft Agent Framework (Microsoft.Agents.AI). Each AIAgent session maps to a long-lived Temporal workflow with full session management: history, StateBag persistence, HITL approval gates, LLM-powered routing, and parallel agent fan-out.

Start here if: you are building with the Microsoft Agent Framework and want durable, stateful, multi-agent sessions.

dotnet add package TemporalCommunity.Extensions.Agents

Full documentation →

For help choosing between the two packages, see the Library Combinations Guide. For the shared approval lifecycle, retry outcomes, and MAF-only scope boundary, see Durable approvals.

How It Works

Both packages share the same core pattern: LLM calls run inside Temporal activities (never directly in workflows), and conversation turns are delivered via Temporal Updates — a durable, acknowledged request/response primitive that eliminates polling.

External Caller
    │
    │  WorkflowUpdate (chat turn / agent message)
    ▼
Temporal Workflow  ←── persists history, serializes turns, handles ContinueAsNew
    │
    │  ExecuteActivityAsync
    ▼
Activity  ←── calls real IChatClient / AIAgent — retried automatically on failure

Prerequisites

• .NET 10 SDK or later to build the repository and run samples
• Temporal Service 1.31.0 or newer (local development: temporal server start-dev)
• An LLM provider (e.g., Azure OpenAI, OpenAI, Ollama)

The repository is tested against Microsoft.Extensions.AI 10.8.3 and Microsoft.Agents.AI 1.17.0 and pins embedded tests to Temporal Server 1.31.2. Package consumers should use those library versions or compatible newer releases and Temporal Service 1.31.0 or newer.

The published libraries ship net10.0 and netstandard2.1 assets. netstandard2.1 supports .NET Core 3.1+ and modern .NET; .NET Framework is out of scope. See each package README for down-level limitations.

Samples

TemporalCommunity.Extensions.AI

TemporalCommunity.Extensions.Agents

Sample credentials

API keys are managed with dotnet user-secrets — stored outside the repo in ~/.microsoft/usersecrets/ and loaded automatically by Host.CreateApplicationBuilder() in the Development environment.

Set OPENAI_API_KEY for each sample project you want to run:

dotnet user-secrets set "OPENAI_API_KEY" "sk-..." --project samples/MEAI/DurableChat

Non-sensitive settings (OPENAI_API_BASE_URL, OPENAI_MODEL, TEMPORAL_ADDRESS) have working defaults in each sample's committed appsettings.json where applicable, or directly in its code. They do not need user-secrets unless you want to override them.

Alternatively, set OPENAI_API_KEY as an environment variable — the samples pick it up automatically via IConfiguration.

# Start Temporal (separate terminal)
temporal server start-dev --namespace default \
  --search-attribute AgentName=Keyword \
  --search-attribute SessionCreatedAt=Datetime \
  --search-attribute TurnCount=Int

# Run a sample
dotnet run --project samples/MEAI/DurableChat
dotnet run --project samples/MAF/BasicAgent

Building

just build        # Restore + Release build
just test-unit    # Unit tests (no server required)
just test         # Unit + integration tests (starts embedded Temporal test servers)
just benchmark-statebag  # Release-mode StateBag rollback timing and allocation measurements
just pack         # Build NuGet packages → artifacts/packages/
just smoke-extensible-turns # Pack, isolate restore, and run the public typed-turn consumer
just ci           # Full pipeline: clean → build → test-unit → pack

License

MIT