Generative AI & agentic solutions · 30–35%

Multi-agent: workflows, Agent Framework, A2A

~3 min read

Foundry workflows (visual designer)

A workflow is a declarative orchestration you draw on a canvas: connected nodes that invoke agents, branch on conditions, loop, and pause for humans — instead of you coding the coordination.

Node types:

Node	Does
Invoke agent	Runs an agent; its structured JSON output can be stored in variables
Flow	If/Else, Go To, For Each (loop lists — e.g. many tickets)
Data transformation	Set/Reset Variable, Parse value
Basic chat	Message user / collect input (human-in-the-loop)
End	Finish, optional result

Patterns: sequential, human-in-the-loop (pause for approval/input), group chat (dynamic control shifts).
Power Fx is Microsoft’s low-code, Excel-like formula language; workflows use it wherever a value or condition is computed: Local.Confidence > 0.8, If(...), ForAll(...), IsBlank(...). System variables describe the conversation; local variables hold your data.
Saved workflows version automatically (immutable versions); visual ↔ YAML stay in sync.
Invoke from code: openai_client.responses.create(conversation=..., extra_body={"agent": {"name": wf_name, "type": "agent_reference"}}).

Microsoft Agent Framework

The Microsoft Agent Framework is an open-source SDK for writing agents and multi-agent orchestrations in code — the next generation of Semantic Kernel + AutoGen, built by the same teams. Unified Agent base class, sessions, function tools (auto schema from type hints or @tool decorator), MCP clients, middleware, telemetry, graph workflows.

Foundry Agent Service = recommended provider — supports service-side chat history (state persists across restarts/scale-out). Azure OpenAI/OpenAI Responses also do; ChatCompletion providers don’t (local history only).
approval_mode on @tool = human confirmation before risky calls.
Agent-as-tool: wrap an agent as a function tool of another agent, so it can delegate work to specialists.
Workflow internals: executors (workers) + edges (direct, conditional, switch-case, fan-out, fan-in) + events (WorkflowOutputEvent, ExecutorInvokeEvent…). Checkpointing = save/resume state.

The 5 orchestration patterns — memorize

Pattern	Shape	Use when	Builder
Concurrent	All agents same task in parallel	Brainstorm, ensemble, voting; independent subtasks	`ConcurrentBuilder`
Sequential	Pipeline: each output feeds the next agent	Fixed ordered steps, progressive refinement	`SequentialBuilder`
Handoff	Full control transfers dynamically	Right expert unknown upfront; escalation/routing	switch-case workflow
Group chat	Shared conversation, manager picks speaker	Debate, maker-checker loops, human oversight (≤3 agents ideal)	`GroupChatBuilder`
Magentic	Manager plans/delegates/adapts + task ledger	Open-ended, no predetermined path, documented plan	`MagenticBuilder`

Match the keyword to the pattern: “parallel / voting” means concurrent; “each step builds on the last” means sequential; “route to the right specialist” means handoff; “maker-checker / debate” means group chat; “open-ended, plan evolves, ledger” means Magentic.

A2A (Agent-to-Agent protocol)

A2A is a standardized protocol that lets agents from different vendors and platforms discover each other and communicate — publish what you can do, receive delegated tasks.

Agent Skill — one capability: id, name, description, tags, examples, I/O modes.
Agent Card — “digital business card”: identity, endpoint URL, capabilities (streaming/push), skills, auth. Served at /.well-known/agent-card.json.
Agent Executor — implements execute (+ optional cancel); uses RequestContext + EventQueue.
Server = Agent Card + request handler (+ Task Store) + ASGI app (Starlette/Uvicorn).
vs MCP: A2A = agent↔agent (each with own LLM, built-in auth); MCP = agent↔tools.