Help Center
DocsCreating Custom Agents
Build your own AI agent with natural language instructions, custom schedules, and configurable permissions
Agent Builder
A 5-step form to create your custom agent
Navigate to AI Agents → + Create Agent to open the builder.
Basics
Instructions
When to Run
Tools & Permissions
Review
Builder Fields Reference
Every configurable field in the agent builder
| Field | Description |
|---|---|
| Name* | Display name shown on agent cards and notifications |
| Description | Short description shown on the agent card (max 500 chars) |
| Instructions* | Natural language instructions telling the agent what to do |
| Trigger Type | How the agent is started: Schedule, Manual, or When something happens (event-based). Event triggers fire when a specific CRM event occurs, such as a deal stage change or new contact creation. |
| Event Conditions | Optional "Only if..." filters for event triggers. Narrow which events actually trigger the agent (e.g., only if deal value > $10k, only if contact source is "Inbound"). |
| Frequency | How often the agent runs (for scheduled triggers) |
| Allowed Tools* | Which CRM tools the agent can use (read and/or write). Per-tool permission overrides let you grant or restrict individual tools beyond the default level. |
| Entity Scope | Automatically determined based on the tools you enable. You do not configure this directly — the system derives which entity types your agent can access from the tools selected in Step 4 (Tools & Permissions). |
| Approval Mode | Whether actions need review before executing: "Ask me first" (all actions go to the review queue) or "Go ahead" (actions auto-execute if confidence meets the threshold) |
| Confidence Threshold | Minimum confidence percentage (10-100%) required for auto-execution in "Go ahead" mode. Actions below this threshold are sent to the review queue instead. Slide left for cautious (more actions queued) or right for aggressive (fewer queued). |
| Per-Tool Overrides | Override the global approval mode for individual write tools. Options: Default (follows global mode), Ask me first (always queue), Go ahead (always auto-execute), Off (disable tool entirely). |
| Max Actions/Run | Safety cap on actions per execution. Defaults to 25 and is configured by your admin — this is not a field in the agent builder form. |
Instruction Templates
5 starter templates to get you going
- Pipeline Review — Flag stale deals and create follow-up tasks
- Follow-up Scheduler — Identify contacts needing outreach
- Data Cleanup — Find duplicates and missing fields
- Lead Qualifier — Score and triage new leads
- Weekly Summary — Generate pipeline health reports
Approval Modes & Confidence
Control how agent actions are handled
| Field | Description |
|---|---|
| Ask me first (recommended) | Every action goes to the review queue. Nothing executes until you approve it. This is the default and safest mode. |
| Go ahead | Actions auto-execute if the AI's confidence meets or exceeds your threshold. Actions below the threshold still go to the review queue. Only available to managers and admins. |
| Smart auto-execute | Automatically executes read operations and low-risk write actions that meet the confidence threshold. Higher-risk actions are queued for review. This mode (auto_low_risk) balances efficiency with safety. |
Event Trigger Reference
Available event types and condition operators
Event types:
- A new company/contact/deal is added (record_created)
- A company/contact/deal is updated (record_updated)
- A deal changes pipeline stage (stage_changed)
Condition operators: is, is not, contains, is empty, is not empty. Up to 5 conditions per trigger.
Available fields depend on the entity type: company (name, type, status, industry, website), contact (first name, last name, email, company, title, status), deal (name, stage, status, value, priority).
Duplicating Agents
Clone an existing agent to create a variation
From an agent's detail page, open the ⋮ menu and select Duplicate. This creates a copy in draft status with "(Copy)" appended to the name. The duplicate includes all settings, instructions, tools, and permissions. Edit the copy to customize it before activating.
Per-Agent Cost Caps
Optional spending limits independent of your org budget
In step 4 of the builder, you can set an optional Daily cost cap and/or Monthly cost capin dollars. These run alongside (and on top of) your organization's overall AI budget — even if the org has plenty of budget left, an individual agent will pause if it hits its own cap.
| Field | Description |
|---|---|
| Daily cost cap | Spending ceiling per UTC day. When hit, the next run fails fast with a "Daily cost budget exhausted" message and an Agent Budget Exhausted notification is sent. Resets at 00:00 UTC. |
| Monthly cost cap | Spending ceiling per UTC calendar month. Same exhaustion behavior as daily; resets on the 1st. |
Agent Chaining
Trigger a second agent automatically when the first completes
On the agent detail page, you can configure a Chain to next agent setting that fires a second active agent on completion of this one. The child run carries a chain_depth counter so multi-step pipelines stay traceable.
Chains never run for dry/test runs, never run if the child is paused or in draft, and silently skip if the child agent already has a run in progress (the existing run continues uninterrupted).
Test Runs (Dry Run on Historical Data)
Validate a draft agent against the last 1–90 days before activating
Click Test on last 7 days in step 4 (or Test Runon an existing agent's detail page) to run the agent against historical data without applying any changes. Test runs:
- Execute the full instructions and tools
- Generate proposed actions marked as preview
- Never write to the CRM
- Are auto-cleaned 25 hours after creation
Retry-After header so the UI can show the cooldown.Versioning
Every save creates a restorable snapshot
Every time you save an agent the previous configuration is captured as a version. From the agent detail page, open the Versions tab to view past configurations and restore any prior version. Restoring is itself a saved change — the version you restore from is preserved.