Action Agents

Action agents run autonomously in the background. Unlike Knowledge agents (which chat with users), Action agents execute on triggers — schedules, webhooks, data changes, or manual runs. They are ideal for ad optimization, data syncing, monitoring, and reporting.

When to use Action Agents

• Ad campaign optimization — Monitor KPIs and adjust budgets, bids, and pacing across Meta, Google, and TikTok Ads
• Spend pacing — Track daily/lifetime spend against targets and take corrective action
• Data monitoring — Watch a URL or API for changes and send alerts
• Scheduled reports — Generate and email daily performance summaries
• CRM sync — Route leads, update records, push notes to HubSpot or Salesforce
• API orchestration — Chain multiple API calls with AI-powered decisions

Creating an Action Agent

1. Go to Agents and click Create Agent
2. Select Action Agent as the type
3. Choose a starter kit (recommended) or start blank
4. Name your agent and proceed to configuration

Starter kits pre-fill the Configure tab with sensible defaults for common use cases — Meta Budget Guard, Google Ads Search Optimizer, Spend Pacing Guard, and more.

Tabs

Action agents have six tabs: Configure, Journey, Triggers, Knowledge, Activity, and Settings.

Configure

The Configure tab is where you define what the agent does and how aggressively it acts. Sections include:

Objective — What the agent optimizes for.

• Type: Threshold (hit a target), Maximize, Rebalance, or Monitor Only
• Target KPI: CPA, ROAS, CTR, CPM, CPC, Spend Pacing, Reach, etc.
• Target value and direction (above/below/within band)
• Evaluation window: 24h, 3 days, or 7 days
• Optional secondary KPIs

Scope — Which campaigns or entities the agent manages.

• Single campaign, selected campaigns, dynamic filter, or all eligible
• Optimization levels: Campaign, Ad Set, Ad, Ad Group
• Auto-enroll new campaigns with optional burn-in period

Budget Control — Spending guardrails.

• Budget level: campaign only, ad set only, or both
• Allow increases/decreases with max percentage caps
• Budget reallocation between entities

Pacing — Spend distribution over time.

• Mode: daily, lifetime, or flight-based
• Target spend per day and utilization percentage
• Tolerance band and check interval (hourly, 4h, daily)
• Overspend action: alert, reduce budget, or pause
• Underspend action: alert, increase budget, or none
• Ramp mode: steady, front-loaded, or back-loaded

Allowed Actions — What the agent can do.

• Toggle individual actions on/off (pause, resume, set budget, etc.)
• Set maximum change percentage per action

Guardrails — Hard limits the agent cannot exceed.

• Min/max budget
• Max actions per run
• Max change percentage
• Cooldown between runs (hours)
• Learning window (days)
• Daily spend cap
• Protected campaigns (excluded from changes)

Approval Mode — How changes get applied.

• Monitor — Agent proposes actions but takes no action; you review in Activity
• Approve First — Agent proposes actions and waits for your approval before executing
• Auto — Agent executes approved action types immediately

Notifications — Get alerts via Slack, email, or in-app on every run, high-impact changes, or failures.

Meta Optimizer Profiles

When the agent is connected to Meta Ads, the Configure tab includes a Meta Optimizer Profile selector. Profiles tune the agent's behavior for specific campaign objectives:

You can also set:

• Variant — Reach, Video Attention, or Impression Efficiency
• Delivery surface — Mixed Meta, Facebook Only, or Instagram Only
• Ownership role — Write Owner, Advisor, or Guard

Google Ads Families

For Google Ads connections, choose an optimizer family:

Google-specific settings include bidding strategy controls (target CPA/ROAS, max CPC), Google Recommendations policy, and launch mode (draft only, validate, approve first, auto apply).

Campaign Ownership

Each campaign can have one write owner — the agent that controls it. This prevents conflicts when multiple agents cover the same ad account. Other agents can be assigned advisor (read-only analysis) or guard (monitoring with alerts) roles.

Journey

The Journey tab uses the same visual flow builder as Knowledge agents, with additional action-specific nodes. Build workflows that fetch data, run AI analysis, propose changes, and execute actions.

Triggers

Each action agent needs at least one trigger:

• Schedule — Cron-based recurring runs (hourly, daily at 9 AM, weekly, etc.)
• Webhook — HTTP POST from external services; payload available as trigger.payload
• Data Monitor — Poll a URL and fire when the response changes (hash comparison)
• Manual — On-demand from the dashboard or API

Knowledge

Link knowledge bases and data tables to give the agent context. For ad optimizer agents, campaign performance data is automatically available from the connected integration.

Activity

Every execution is logged with:

• Status: Pending, Running, Completed, Failed
• Trigger: Which trigger started the run
• Duration: How long the run took
• Proposed actions: What the agent wants to do, with risk levels and reasoning
• Execution results: What was actually applied

In Approve First mode, pending proposals appear here for review. Approve or reject individual actions.

Settings

• General — Name, description, status
• Groups — Add the agent to a group for multi-agent coordination
• Danger Zone — Delete the agent

Available Kits

Starter kits pre-configure the agent for specific use cases:

Publishing

When you publish an action agent:

• Schedule triggers become active immediately
• Webhook URLs start accepting requests
• Data monitors begin polling

Unpublishing stops all triggers and active schedules.

Multi-agent Coordination

Action agents can work together:

• Agent Groups — Add agents to a group (in Settings) to share context. Knowledge agents in the same group can reference action agent run results.
• Trigger Agent node — One agent can trigger another from its workflow, passing input variables and optionally waiting for results.
• Chaining — Build pipelines where multiple agents run in sequence, each handling a different part of the workflow.