6.1 Add Agent Flows to a Topic

6.1 Add Agent Flows to a Topic

Quick Answer: Build an agent flow with When an agent calls the flow and Respond to the agent (async off). Publish it, then either register it as an agent tool for generative orchestration or insert it as an action node inside a topic. Bind inputs from variables or AI slot-filling, capture outputs, then continue the topic with messages, conditions, or further nodes.

AB-620 treats topics as the conversation control surface and agent flows as the multi-step automation engine. Domain 1 leaf skill Add agent flows to a topic expects you to wire those two surfaces together correctly—not merely create a flow that never returns data to the agent.

Why topics need agent flows

A topic alone can ask questions, set variables, send messages, and call many tools. An agent flow shines when you need:

• Multi-step connector choreography (read Dataverse, create ServiceNow incident, email confirmation)
• Human-in-the-loop approvals or long-running steps designed after the agent response
• Consistent, deterministic automation that should not depend on free-form generative planning for every micro-step
• Secrets or shared connections managed with flow run-only settings rather than ad-hoc topic HTTP calls

Microsoft positions agent flows as native Copilot Studio automations (Power Automate–like designer, optimized for agents). You run them as standalone automations or attach them so an agent can invoke them.

Flow prerequisites the exam loves

Before a topic can call a flow as a tool, the flow must meet published requirements:

When you create a new agent flow from the Flows page or as a new tool on the agent, Copilot Studio scaffolds the required trigger and response action. Converting an older cloud flow means swapping in that trigger and response pattern.

Supported input types and hygiene

Runtime guidance commonly expects clean base types—text, boolean, numbers—and non-null values for required inputs. Complex nested objects often need flattening into text or numbers the agent can pass. Null required inputs are a top cause of silent failures at the action node.

Two placement models: agent tool vs topic action

Microsoft documents two mechanisms that both matter on AB-620:

1) Add the flow as an agent-level tool

Path (conceptual):

1. Open the agent → Tools → Add a tool → Flow.
2. Select a published flow → Add and configure.
3. Set Name and a clear Description for the orchestrator.
4. Configure each Input fill mode (dynamic AI fill vs custom value).
5. Configure Completion (how results return into the conversation).
6. Save, then test in the agent test panel.

With generative orchestration (default for new agents), the planner can choose this flow tool based on name, description, inputs/outputs, and conversation context—without a hard-coded topic path—when Allow agent to decide dynamically when to use the tool is enabled.

2) Call the flow explicitly from inside a topic

Path (conceptual):

1. Open Topics → edit or create the topic (trigger phrases or generative description as appropriate).
2. Add node → Add a tool / call a tool, then select the flow (or create new flow from the node menu where available).
3. Map each flow input to a variable, literal, or Power Fx formula.
4. On success, map output parameters into topic or conversation variables.
5. Continue with Message, Condition, Question, adaptive card, or redirect nodes that consume those outputs.
6. Save the topic and republish the agent after meaningful changes.

Classic orchestration agents only call tools (including flows) when a topic node invokes them. Generative orchestration can still benefit from explicit topic calls when you need a fixed sequence—collect card inputs, validate, call flow, then format a compliance-approved message.

Input and output binding patterns

Binding is where exam scenarios hide traps.

Filling flow inputs

• From topic variables: User answered a Question node or submitted an Ask with Adaptive Card; map those variables into trigger inputs.
• From conversation context / generative slot filling: At agent-tool level, default Dynamically fill with AI lets the orchestrator extract values or generate clarifying questions—similar to generative topic input filling.
• Override with custom value: Set a fixed literal, an existing variable, or Power Fx when the value must not be guessed (for example environment-specific IDs from environment variables, or a known case number already captured).

Returning outputs to the topic

The Respond to the agent action defines output parameters. After the action node runs:

1. Capture each output into a topic variable.
2. Use those variables in a Message node (including formatting and adaptive content).
3. Branch with Condition nodes (for example status = Failed → escalation topic).
4. Pass outputs into another tool or a redirected topic’s input parameters.

Tip aligned with generative design: for agents using generative orchestration, you can return structured outputs from a topic and let the agent summarize them instead of always hard-coding the final user-facing sentence—while still using flows for the deterministic system writes.

Classic vs generative orchestration implications

Exam trap: Enabling generative orchestration does not remove the need for explicit topic calls when compliance requires a fixed order (identity check → adaptive card → flow → confirmation card). Generative flexibility and deterministic topics coexist; choose by risk, not fashion.

End-to-end scenario (device request)

Contoso IT builds Request device:

1. Topic trigger description: “Helps employees request a replacement laptop or phone and creates a tracked ticket.”
2. Ask with Adaptive Card collects device type, justification, and cost center into variables.
3. Condition validates cost center format; on failure, message + re-ask.
4. Action node calls agent flow CreateDeviceRequest with inputs DeviceType, Justification, CostCenter, UserEmail (from system user context where available).
5. Flow writes Dataverse row, creates ServiceNow request, returns TicketNumber and Status.
6. Message node: “Your request {TicketNumber} is {Status}. IT will follow up within one business day.”
7. Optional quick replies: “Check request status,” “Talk to an agent.”

If the same flow is also agent-level with a vague description “Handles IT stuff,” generative orchestration may fire it on unrelated IT questions. Tighten the description and consider disabling dynamic use when the flow must only run after card validation.

Troubleshooting checklist

1. Open the topic action node → refresh parameters if the flow schema changed.
2. Confirm Copilot Studio inputs/outputs match the flow trigger and Respond action names/types.
3. Save topic, republish agent, and retest a fresh conversation.
4. In the flow, verify Asynchronous response is Off on Respond to the agent.
5. Reject null required inputs; stick to supported base types.
6. Use the flow checker and agent activity map (generative) to see whether the planner selected the tool or the topic path ran.
7. Confirm connections: end-user credentials vs maker-provided connections for run-only users, especially in CMK environments where user-provided connections may be restricted.

Design decisions for AB-620 stems

• Prefer topic + flow when the user must complete a structured multi-field form before any system write.
• Prefer agent-level flow tool when the action is a single capability the planner should discover from natural language.
• Prefer human-in-the-loop flow patterns when high-risk changes need approval (covered in the agent-flows chapters)—still return a clear status to the agent after the gate.
• Never assume “add flow somewhere” without Respond to the agent outputs the conversation can use.

Success checklist

• Flow has agent trigger + Respond action, async off, published.
• Inputs bound from variables or deliberate AI fill; no silent nulls.
• Outputs stored and used in messages or branches.
• Placement matches orchestration mode and compliance needs.
• Descriptions prevent accidental generative double-calls.
• Tested after every schema change with republish.

Mastering this skill connects conversation UX to enterprise automation—the core of “configure topics” on AB-620 Domain 1.