Agents
Open-ended decisions, conversation, and autonomous tool use
Choose the model, session, context providers, tools, middleware, MCP clients, and approval policy.
Independent developer guide · August 3, 2026
Microsoft Agent Framework is an open-source SDK for agents and multi-agent workflows. Its agents, Harness, and workflows share foundations, but they solve different control problems—and none of them automatically grants safe data or tool authority.
Architecture decision
Open-ended decisions, conversation, and autonomous tool use
Choose the model, session, context providers, tools, middleware, MCP clients, and approval policy.
One agent carrying a long task across many service calls
Stable Python and .NET core; verify the separate lifecycle and permissions of optional file, shell, loop, and background features.
Known graphs with routing, checkpoints, functions, agents, and human review
Use programmatic graphs for custom logic or declarative YAML for standard, reviewable sequences.
A deterministic operation with known inputs and outputs
Prefer ordinary code when it is sufficient; an agent adds cost, uncertainty, and a new authority surface.
SDK support
Install from the reviewed Agent Framework packages and pin the exact set used by agents, Harness, workflows, and providers.
Current Learn examples can use prerelease package flags even where a component has a stable announcement; verify each artifact.
The overview currently excludes declarative agents, RAG, CodeAct, and functional workflows from Go support.
Package feeds, provider integrations, adapters, and optional capabilities can have narrower lifecycle labels than the core SDK. Lock and test the exact dependency graph.
Production path
01
Choose a function, workflow, agent, Harness, or Squad from the task shape—not from feature breadth.
02
Pin SDK, provider, adapter, model, and schema versions; record preview flags and runtime warnings.
03
Separate data access from action authority, and scope file, shell, URL, MCP, and state-changing tools independently.
04
Set hard budgets for calls, tokens, elapsed time, loops, retries, and external spend outside model instructions.
05
Evaluate fixed cases, adversarial inputs, denied tools, retries, cancellation, resume, and target-system receipts.
Implementation guides
Configure the current Python and .NET entry points, compaction, approvals, optional tools, isolation, and tracing.
Compare the current Python and .NET YAML shapes, actions, handlers, human review, and checkpoint tests.
Connect the provider and team adapter without copying approve-all permissions into a normal environment.
Choose one long-task runtime or a coordinator with persistent specialists, then bound ownership and authority.
This page is an independent information guide, not an embedded Microsoft runtime or a Flowith integration. Source check: August 3, 2026. Verify current SDKs, packages, providers, lifecycle labels, model support, storage, tools, approvals, telemetry, third-party terms, and deployment controls.