AI Agents
AI agents in Outrun are workflow-powered automation that can chat with visitors, triage email, update your CRM, and route conversations to specialist agents. Each agent is backed by a workflow template that defines its trigger, AI reasoning, tool use, and response logic.
Agent Templates
Outrun includes pre-built agent templates for common use cases. Each template creates a complete workflow with triggers, AI nodes, routing, and actions pre-configured.
| Template | Channel | Description |
|---|---|---|
| Pre-Sales Agent | Chat | Answers product questions, explains pricing, qualifies interest |
| Lead Qualification Agent | Chat | Asks qualifying questions, captures contact details, creates CRM records |
| Customer Success Agent | Chat | Handles account questions, troubleshooting, and feature guidance |
| AI SDR | Drafts outbound sales emails using CRM context | |
| Email Drafter | Generates email responses based on inbound message analysis | |
| Activity Digest | Schedule | Summarises daily activity across connected tools |
| Data Enricher | Schedule | Fills in missing fields on CRM records using AI analysis |
| Duplicate Detector | Schedule | Identifies and flags duplicate records across sources |
| Smart Responder | AI-powered email responses with tone and context matching | |
| Deal Analyzer | CRM | Analyses deal health and recommends next actions (HubSpot) |
| Lead Scorer | CRM | Scores incoming leads based on configurable criteria (HubSpot) |
| Ticket Router | Webhook | Routes inbound tickets to the right team (Jira) |
Creating an Agent
- Navigate to Agents in the sidebar
- Click New Agent
- Choose a template or start from scratch
- Configure the agent settings (see below)
- Activate the agent
When you create an agent from a template, Outrun generates a workflow with the correct node structure. You can customise the workflow in the canvas editor.
Agent Settings
Identity
- Agent Name — displayed in chat widget and email headers (e.g. "Sarah from Acme")
- System Prompt — instructions that define the agent's personality, knowledge, and boundaries
- Model — the AI model used for reasoning (Claude Haiku for speed, Claude Sonnet for quality)
Channels
Agents receive messages through one or more channels:
- Chat — embeddable website widget via the Outrun SDK
- Email — connected email integrations (inbound messages trigger the agent)
- Webhook — external events from GitHub, Jira, or custom sources
- Schedule — cron-based triggers for periodic tasks
Configure channels in the Channels tab of the agent settings. Chat agents require the SDK to be installed on your website.
Tool Use
Agents can dynamically choose which actions to take during a conversation. Available tools include:
| Tool | Description |
|---|---|
crm_lookup_contact |
Search for a contact in your connected CRM |
crm_update_deal |
Update deal stage, value, or custom fields |
send_email |
Send an email via connected email integration |
fetch_workspace_context |
Pull workspace data for context-aware responses |
create_contact |
Create a new CRM contact record |
Tools are registered in the workflow's AI node configuration. The agent decides which tools to call based on conversation context — it reasons about what action would be most helpful, calls the tool, and uses the result in its response.
Tool Use vs. Scripted Actions
Traditional workflow actions run in a fixed order. Tool use lets the agent decide what to do based on the conversation. The agent might look up a contact, realise they're on the enterprise plan, and tailor its response accordingly — without you needing to pre-define that branch.
Human-in-the-Loop (HITL)
HITL settings control when agent responses require human review before being sent:
- Mode —
review(all responses reviewed),confidence(low-confidence responses reviewed), orautonomous(agent responds directly) - Blocking — when enabled, the agent waits for human approval before responding. When disabled, the agent responds immediately and corrections are applied after the fact.
- Email Capture Delay — how long (in seconds) the agent waits before asking for the visitor's email address. Configurable per agent.
Review pending responses in the HITL Review page, which shows the agent's draft alongside the full conversation context.
Multi-Agent Routing
When you deploy multiple chat agents, Outrun automatically creates a triage workflow that routes incoming messages to the right specialist.
How Routing Works
- A visitor sends a message via the chat widget
- The auto-router (a lightweight triage agent) reads the message and classifies it
- The message is routed to the specialist agent that matches the classification
- The specialist responds using its own prompt, tools, and context
- If the conversation topic changes, the auto-router can re-route to a different agent
Routing Configuration
The auto-router uses conditional nodes to match triage classifications to specialist agents. Each specialist agent has a sourceMode and agentName that the router uses for targeting.
Routing decisions are logged in the workflow run history so you can audit which agent handled which conversation and why.
Loop Guards
To prevent infinite routing loops (Agent A routes to Agent B, which routes back to Agent A):
- A hop counter tracks how many times a conversation has been re-routed
- After 3 hops, the conversation escalates to the HITL queue
- An agent cannot route back to the agent that just routed to it
Correction Memory
When a human edits an agent's response in the HITL review queue, the correction is stored and used to improve future responses.
How It Works
- The agent generates a draft response
- A human reviews and edits the response
- The original message, draft, and correction are stored as a correction record
- When a similar message arrives in the future, relevant corrections are retrieved
- The corrections are injected into the agent's prompt as few-shot examples
Viewing Corrections
Navigate to the agent's Correction Memory Trail in settings to see stored corrections, how often they've been retrieved, and their impact on response accuracy.
When to Update the Prompt Instead
If you've stored more than 5 corrections for the same underlying issue, it's more effective to update the system prompt directly. Correction memory handles edge cases; systemic issues belong in the prompt.
