Autonomous Agents

Agents are automated AI loops that perform tasks on your connected sites without manual intervention. Unlike chat (where you send one message at a time), an agent runs a multi-round loop: it sends a prompt to the AI, executes tool calls, feeds results back, and repeats until the task is complete or the round limit is reached.

What is an Agent

An agent consists of:

• Name — a descriptive identifier
• System prompt — instructions that define what the agent should do
• Model — the AI model to use (e.g., Claude Sonnet 4)
• Sites — which connected sites the agent can interact with
• Max rounds — the maximum number of AI-tool interaction loops (default: 5)
• Trigger type — how the agent is started (manual, scheduled, or webhook)
• API key — which AI key to use (platform credits or a specific BYOK key)

Creating an Agent

1. Navigate to Agents in the sidebar.
2. Click New Agent or Start from Template.
3. Fill in the configuration:
   • Enter a name and optional description
   • Write the system prompt (the core instructions for the AI)
   • Select the model and sites
   • Set the max rounds slider (1-20)
   • Choose the trigger type
4. Click Create.

Agent Templates

The template gallery offers pre-built agents for common scenarios:

• Content Monitor — watches a site for new content
• Forum Moderator — reviews new posts against community guidelines
• Data Summarizer — aggregates data from a site into a digest
• Support Responder — drafts replies to support threads

Select a template, customize the prompt and sites, and save.

Trigger Types

Manual

Run the agent on demand by clicking Run Now on the Agents page. Suitable for one-off tasks or testing.

Scheduled (Cron)

Set a schedule using cron expressions or presets:

Preset	Cron Expression	Runs
Every hour	0 * * * *	Top of every hour
Daily at 09:00	0 9 * * *	Once per day
Every Sunday at 09:00	0 9 * * 0	Once per week
First of month at 09:00	0 9 1 * *	Once per month
Every 15 minutes	*/15 * * * *	4 times per hour

For custom schedules, enter a cron expression directly. The next run time is displayed for confirmation.

INFO

Scheduled agents require the Harness plan or higher.

Webhook

Trigger the agent via an HTTP POST request from an external system:

1. Select Webhook as the trigger type.
2. Save the agent to receive the webhook URL and secret.
3. Configure your external system to POST to the webhook URL.

The webhook payload can include dynamic data that gets injected into the agent's context. See the Webhook Triggers guide for payload format and authentication.

INFO

Webhook-triggered agents require the Harness plan or higher.

Running and Monitoring

Starting a Run

• Manual: Click Run Now on the agent card
• Scheduled: Runs automatically at the scheduled time
• Webhook: Runs when the webhook URL receives a POST request

Run Timeline

Each run produces a timeline of steps:

Step Type	Description
Thinking	The AI's reasoning before a tool call
Tool Call	A tool invocation with parameters
Tool Result	The response from the site
Output	The AI's final text output
Error	An error that occurred during the step

Navigate to Agents > [Agent] > Runs to see all runs. Click a run to see the full timeline.

Live Monitoring

Active runs show a real-time view. Steps appear as they execute, and you can watch the agent work through its task.

Cancelling a Run

Click Cancel Run on an active run to stop it. The agent finishes the current step and halts. Credits consumed up to that point are still deducted.

Agent Run Statuses

Status	Meaning
Pending	Queued, waiting for a worker
Running	Actively executing
Done	Completed successfully
Failed	Encountered an error
Cancelled	Stopped by the user

Credit Usage

Each round of an agent run consumes credits based on the model and token count, just like chat messages. The total credit cost is shown on the run detail page.

To control costs:

• Set a conservative maxRounds value
• Use cheaper models for simple monitoring tasks
• Use BYOK keys for high-frequency agents

Safety

Principle of Least Privilege

Give agents access only to the sites and tools they need. Use site groups with deny lists to block destructive operations.

Human-in-the-Loop

For critical actions, design your agent prompt to output a plan rather than execute directly. Review the plan in the run timeline before creating a follow-up agent that executes.

Max Rounds as a Safety Net

Set maxRounds to the minimum needed for the task. A content monitoring agent might need only 2-3 rounds, while a complex multi-site workflow might need 10+.

Workspace Agents

In a Team workspace, agents can be shared among workspace members. Workspace agents use the workspace's shared vault tokens and can be triggered by any member with the appropriate permissions. See Team Workspaces for details.