# Agents

Agent components use an LLM to reason and act. They range from general-purpose autonomous agents to specialised routing and delivery agents.

---

## Agent

A general-purpose autonomous agent that uses an LLM with tools and knowledge bases to reason and act over multiple steps.

**Use in a flow**

1. Add an **Agent** node to the canvas.
2. Select **Model Provider** and **Model ID**.
3. Write **Agent Instructions** (system prompt).
4. Optionally connect **Tools**, select a **Knowledge Base**, and wire **Input** from **Input** or upstream components.
5. Connect **Selected Output** to **Output** or the next step.

```
Input ──► Agent ◄── Tools / MCP Client
Knowledge Base (dropdown) ──► Agent
Agent (Selected Output) ──► Output
```

The agent can call connected tools, retrieve knowledge base context, use memory, and run guardrails before returning a response.

**Parameters — Model and instructions**

| Parameter | Default | Hidden | Description |
| --- | --- | --- | --- |
| Model Provider | — | No | LLM provider for the agent. |
| Model ID | — | No | Model to use. Refresh to load available models. |
| Region Name | — | No | Region where the selected model is available. |
| Agent Instructions | — | No | System prompt that guides agent behavior. |
| Input | — | No | User message or task for the agent to process. |

**Parameters — Tools and knowledge**

| Parameter | Default | Hidden | Description |
| --- | --- | --- | --- |
| Tools | — | No | Connected tool components the agent can call. |
| Knowledge Base | — | No | Optional knowledge base for retrieval-augmented responses. Refresh to load options. |
| Fallback Knowledge Base | — | Yes | KB used when the user lacks access to the primary KB. |
| Assume Policy | — | No | Access policy limiting which resources this agent can use. |

**Parameters — Memory and behavior**

| Parameter | Default | Hidden | Description |
| --- | --- | --- | --- |
| STM Retrieval Limit | 10 | Yes | Recent conversation turns from short-term memory. |
| LTM Retrieval Limit | 10 | Yes | Records from long-term memory via semantic search. |
| User Level LTM | false | Yes | Scope long-term memory per user instead of shared agent namespace. |
| Current Date | false | Yes | Add a tool that returns the current date. |
| Max Iterations | 15 | Yes | Maximum reasoning steps before the agent stops. |
| Handle Parse Errors | true | Yes | Recover from input parsing errors. |
| Verbose | true | Yes | Detailed agent logging during execution. |

**Parameters — Delivery and personalization**

| Parameter | Default | Hidden | Description |
| --- | --- | --- | --- |
| Enable File Deliverables | false | No | Let the agent register files for delivery to channels. |
| Enable Channel Delivery | false | No | Inform the agent that a downstream Channel Agent handles outbound delivery. |
| Post-HITL Instructions | — | Yes | Custom instructions after Human In The Loop approval. |
| Enable Personalization | false | No | Inject user personalization preferences into the prompt. |

**Parameters — Structured output**

| Parameter | Default | Hidden | Description |
| --- | --- | --- | --- |
| Enable Structured Output | false | No | Return JSON matching a defined schema. |
| Output Schema | — | No | Table of field names, types, and descriptions. Appears when structured output is enabled. |
| Output Field | response | No | Which field to expose on **Selected Output** when structured output is enabled. |
| Schema Name | OutputModel | Yes | Name for the output schema model. |
| Format Instructions | — | Yes | Instructions for formatting structured output. |

**Parameters — Guardrails**

| Parameter | Default | Hidden | Description |
| --- | --- | --- | --- |
| Enable Input Guardrail | false | No | Validate input before agent execution. |
| Input Guardrails | — | Yes | Guardrails to apply on input. Appears when enabled. |
| Enable Output Guardrail | false | No | Validate agent output after execution. |
| Output Guardrails | — | Yes | Guardrails to apply on output. Appears when enabled. |

**Outputs**

| Output | Description |
| --- | --- |
| Selected Output | Agent response message, or a specific structured output field. |
| Language Model | The built language model instance (for wiring to other components). |

**Limitations**

- **Model required** — A model must be selected before the agent can run.
- **Tool mode** — When used as a tool by another agent, the deprecated **Agent Description** applies.
- **Knowledge base** — Retrieval uses the selected KB’s embedding model and search settings internally.

---

## Intent Classification Agent *(Beta)*

Classifies the user’s intent and returns a structured output with the classified category.

**Use in a flow**

1. Add an **Intent Classification Agent (Beta)** node to the canvas.
2. Select **Model Provider** and **Model ID**.
3. Define **Agent Instructions** and an **Output Schema** (for example, an `intent` string field).
4. Connect **Input** from **Input** or an upstream message.
5. Connect **Structured Output** downstream (for example, a conditional router or **Agent**).

```
Input ──► Intent Classification Agent ──► Router / Agent
```

**Parameters**

| Parameter | Default | Hidden | Description |
| --- | --- | --- | --- |
| Model Provider | — | No | LLM provider for classification. |
| Model ID | — | No | Model to use. |
| Region Name | — | No | Region where the model is available. |
| Agent Instructions | helpful classifier prompt | No | System prompt guiding intent classification. |
| Input | — | No | User query to classify. |
| Output Schema | intent field | No | Fields to return (name, description, type). |
| Output Field | structured_output | No | `structured_output` for full JSON, or a specific schema field name. |
| Enable Streaming | false | No | Run classification in streaming mode. |
| Schema Name | — | Yes | Name for the output schema model. |
| Model Kwargs | — | Yes | Extra keyword arguments for the model. |

**Outputs**

| Output | Description |
| --- | --- |
| Structured Output | Classified intent as a message. Returns full JSON or a single field value based on **Output Field**. |

**Limitations**

- **Beta** — Behavior may change.
- **Schema required** — **Output Schema** must define at least one field.
- **Structured output** — Uses native structured output when the model supports it; otherwise falls back to prompt-based parsing.

---

## Human In The Loop *(Beta)*

Pauses the flow and sends an approval request to configured delivery targets. The flow resumes once approved or rejected.

**Use in a flow**

1. Add a **Human In The Loop (Beta)** node after an upstream **Agent** (or any component that produces a bot response).
2. Wire **User Query** and **Bot Response** from upstream steps.
3. Configure **Approval Delivery Targets** (channel, recipient, event type).
4. Connect **Response** downstream.

```
Input ──► Agent ──► Human In The Loop ──► Channel Agent / Output
```

On first run, the component sends an approval card to the configured channels and pauses execution. When a human approves or rejects, the flow resumes with the action result.

**Parameters**

| Parameter | Default | Hidden | Description |
| --- | --- | --- | --- |
| User Query | — | No | Original user message for approval context. |
| Bot Response | — | No | Generated answer included in the approval request. Required. |
| Approval Title | Action Required | No | Title on the approval card. |
| Approval Description | default text | No | Description explaining what needs approval. |
| Approval Delivery Targets | — | No | Where to send approval requests (channel, event type, recipient). One row. Refresh to load routes. |
| Send Reply to Original Channel | true | No | Send confirmation back to the incoming channel after dispatching the approval. |
| Attach Incoming Files | false | No | Attach session-uploaded files to the approval card. |
| Post-Approval Instructions | — | No | Custom instructions injected into the main agent after approval. |
| Timeout (seconds) | 86400 | No | Timeout value in seconds. Under advanced settings. |
| Auto-reject on Timeout | false | No | Under advanced settings. |
| Context Data (JSON) | {} | No | Extra context shown on the approval card. Under advanced settings. |
| Assume Policy | — | No | Access policy for channel and resource access. |

**Outputs**

| Output | Description |
| --- | --- |
| Response | On pause: JSON with `awaiting_approval` status. On resume: approval or rejection message. |

**Limitations**

- **Beta** — Behavior may change.
- **Delivery target required** — One complete row in **Approval Delivery Targets** (channel, event type, recipient). Approval requests are delivered to Microsoft Teams.
- **Playground and NeoGPT** — Approval requests are not raised in these channels; the flow continues without review.
- **Flow pause** — Execution stops until a human responds.
- **Event type** — Approval targets use the `hitl_approval` event type automatically.

---

## Channel Agent *(Beta)*

An agent wired to a communication channel. Handles inbound messages and routes outbound replies to configured delivery targets.

**Use in a flow**

1. Add a **Channel Agent (Beta)** node after an upstream **Agent** that generates the message content.
2. Wire **User Query** and **Bot Response**.
3. Configure **Outbound Routes** or **Delivery Targets**.
4. Connect **Response** downstream.

```
Input ──► Agent ──► Channel Agent ──► Output
```

The Channel Agent resolves the correct channel and recipient from the user query and delivers the bot response. Fully configured delivery targets are sent automatically; partial rows are resolved by the LLM at runtime.

**Parameters**

| Parameter | Default | Hidden | Description |
| --- | --- | --- | --- |
| User Query | — | No | Original user query — used to determine channel, recipient, and event type. |
| Bot Response | — | No | Content from an upstream agent to deliver. Required. |
| Model Provider | — | No | LLM provider for routing decisions. |
| Model ID | — | No | Model used to resolve delivery targets. |
| Agent Instructions | auto | No | Base system prompt. `auto` generates routing instructions from configured routes. |
| Additional Instructions | — | No | Extra routing or delivery overrides appended to the prompt. |
| Outbound Routes | — | No | Channels this agent can deliver to. Refresh to load routes. |
| Delivery Targets | — | No | Table of outbound targets (channel, event type, recipient, invocation). Empty fields are resolved by the LLM. |
| Send Reply to Original Channel | true | No | Send confirmation back to the incoming channel after delivery. |
| Max Iterations | 10 | Yes | Maximum agent reasoning steps. |
| Handle Parsing Errors | true | Yes | Recover from tool parsing errors. |
| Verbose | false | Yes | Detailed agent logging. |
| Model Kwargs | — | Yes | Extra keyword arguments for the model. |

**Outputs**

| Output | Description |
| --- | --- |
| Response | Delivery result message from the channel agent. |

**Limitations**

- **Beta** — Behavior may change.
- **Routes required** — At least one **Outbound Route** or **Delivery Target** must be configured.
- **Bot response required** — Upstream content must be wired to **Bot Response**.
- **Automatic targets** — Targets marked automatic must have channel, event type, and recipient fully configured.

---

## Flow Agent

Calls another Agentic Flow as a sub-flow and returns its output into the current flow.

**Use in a flow**

1. Add a **Flow Agent** node to the canvas.
2. Select **Workspace** and **Flow Agent** (refresh to load options).
3. Wire **Input** from an upstream message or component.
4. Optionally set **Tweaks** to override sub-flow parameters.
5. Connect **Response** or **Full Response** downstream.

```
Input ──► Flow Agent ──► Output / Agent
Agent ──► Flow Agent (as a tool)
```

**Parameters**

| Parameter | Default | Hidden | Description |
| --- | --- | --- | --- |
| Input | — | No | Input message or data for the sub-flow. |
| Workspace | — | No | Workspace containing the target flow. Refresh to load. |
| Flow Agent | — | No | Agentic flow to execute. Refresh after selecting a workspace. |
| Conversation history | — | No | Optional conversation history passed to the sub-flow. |
| Session ID | — | No | Session ID for conversation continuity. |
| Output Component | — | No | Specific output component to retrieve from the sub-flow. |
| Tweaks | — | No | Key-value parameter overrides for the sub-flow. Auto-populated from the selected flow. |
| Enable Retry | — | Yes | Retry failed API requests. |
| Max Retries | — | Yes | Maximum retry attempts. |
| Save Output as Variable | — | Yes | Store output as a named flow variable. |
| Variable Name | — | No | Name for the flow state variable. |

**Outputs**

| Output | Description |
| --- | --- |
| Response | Main text output from the sub-flow execution. |
| Full Response | Complete API response including metadata and all output data. |

**Limitations**

- **Workspace and flow required** — Both must be selected before execution.
- **Authorization** — Runs using the flow’s authorization context; sub-flow must be accessible to the caller.
- **Session continuity** — Pass **Session ID** to maintain conversation context across parent and child flows.
- **Tweaks** — Values support `{{variable}}` placeholders resolved from flow context.

---

## Voice Agent Call

Triggers a Voice Agent call and returns the call response.

**Use in a flow**

1. Add a **Voice Agent Call** node to the canvas.
2. Select **Workspace** and **Voice Agent** (refresh to load options).
3. Fill in **Call Parameters** (auto-populated when a voice agent is selected).
4. Connect **Call Response** or **Status** downstream.

```
Flow trigger ──► Voice Agent Call ──► Output / Agent
```

**Parameters**

| Parameter | Default | Hidden | Description |
| --- | --- | --- | --- |
| Workspace | — | No | Workspace containing the voice agent. Refresh to load. |
| Voice Agent | — | No | Voice agent to trigger. Refresh after selecting a workspace. |
| Call Parameters | — | No | Parameter table auto-populated from the selected voice agent. Fill in values for each parameter. |
| Enable Retry | — | Yes | Retry failed API requests. |
| Max Retries | — | Yes | Maximum retry attempts. |
| Save Output as Variable | — | Yes | Store output as a named flow variable. |
| Variable Name | — | No | Name for the flow state variable. |

**Outputs**

| Output | Description |
| --- | --- |
| Call Response | API response from the voice call (result, status code, raw data). |
| Status | HTTP status code from the last call attempt. |

**Limitations**

- **Workspace and voice agent required** — Both must be selected before triggering a call.
- **Call parameters** — Required parameters depend on the selected voice agent configuration.
- **Authorization** — Requires valid authorization to call the voice backend service.

---

## Skills

Resolves workspace skills and outputs them for an **Agent** to use. See [Skills](../../12_foundry/integrations/skills.html) for creating and managing skills.

**Use in a flow**

1. Add a **Skills** node and select one or more active skills.
2. Set each skill to **Allow** or **Ask** in **Skill Approvals**.
3. Connect **Skills** to the **Skills** input of an **Agent**.

```
Skills ──► Agent ──► Output
```

The agent receives a short list of the selected skills and reads a skill’s `SKILL.md` only when a request matches it.

**Parameters**

| Parameter | Default | Hidden | Description |
| --- | --- | --- | --- |
| Workspace ID | — | Yes | Workspace used to fetch available skills. Filled automatically. |
| Skill Provider | AGENTCORE | No | Runtime provider for downloading and hosting skills. |
| Skills | — | No | Skills to load into the agent sandbox. Refresh to load the workspace’s active skills. |
| Sandbox Status | Not checked | No | Refresh to test connectivity to the sandbox runtime. |
| Skill Approvals | Allow | No | One row per selected skill. Set a row to **Ask** to pause for approval before the agent uses that skill. Rows follow the **Skills** selection and cannot be added or deleted. |

**Outputs**

| Output | Description |
| --- | --- |
| Skills | The selected skills, for the Agent’s **Skills** input. |

**Limitations**

- **Approval destination** — A skill set to **Ask** sends its approval request to the delivery target of a **Human In The Loop** node in the flow.
- **Multiple nodes** — Several Skills nodes can be connected to one Agent.

---

## Break Agent

Immediately terminates the current flow execution with an optional break reason and instruction.

**Use in a flow**

1. Add a **Break Agent** node where you want to pause execution after processing.
2. Configure it like a standard **Agent** (model, instructions, tools, input).
3. Optionally set **Break Instruction** and **Break Reason**.
4. Connect **Selected Output** to **Output** so the response is delivered before the pause.

```
Input ──► Agent steps ──► Break Agent ──► Output
```

The agent processes the input, streams the response downstream, then pauses the flow. On resume, it runs as a normal agent without breaking again.

**Parameters**

Break Agent includes all [Agent](#agent) parameters plus the following:

| Parameter | Default | Hidden | Description |
| --- | --- | --- | --- |
| Break Instruction | execution state summary prompt | No | Instruction injected when breaking. The agent summarizes execution state. |
| Should Break | true | Yes | When `false`, behaves like a normal Agent without pausing. |
| Break Reason | agent_break | Yes | Reason logged for the break (for debugging). |

**Outputs**

| Output | Description |
| --- | --- |
| Selected Output | Agent response message before the break is set. |

**Limitations**

- **Pause after response** — The response is sent to downstream components before execution pauses.
- **Resume behavior** — On resume (`is_resume` in flow context), the agent does not break again.
- **Execution ID** — Flow can be resumed later using the execution ID stored in flow context.
- **Should Break off** — Set **Should Break** to `false` to use as a standard Agent without pausing.