Docs

Custom agents

Create custom agents for repeatable Workbase workflows.

Custom agents are for repeatable workflows. They should have one clear job, connected context, visible ownership, and run history that can be reviewed.

Good use cases

  • Inbound Qualification: Review new inbound companies, check fit, summarize buyer context, and prepare a handoff.
  • Pipeline Follow-up: Review stale deals, identify the last meaningful interaction, and create follow-up tasks.
  • Account Research: Research target accounts before outreach and enrich the company record.
  • Deliverability Review: Review bounced outreach, group failure reasons, and recommend cleanup actions.

Before you create one

Make sure the workflow already has a clear source of truth. For example, inbound qualification should have inbox conversations and company records. Pipeline follow-up should have deal records or a board. Account research should have target account records or a list.

Agent configuration checklist

  1. Name the job

    Use a specific name that describes the workflow, such as Inbound Qualification or Pipeline Follow-up.

  2. Set ownership

    Set an owner or team so people know who maintains the agent.

  3. Write the prompt

    Explain the task, the expected output, and when the agent should ask for approval.

  4. Connect context

    Add the resources the agent needs, such as an inbox channel, object, integration, MCP tool, or app.

  5. Review first runs

    Run the agent on a small sample before relying on it for a full workflow.

Prompt guidance

Good prompts are operational. They tell the agent what to inspect, what output to produce, what not to do, and when to request human review.

Avoid prompts like “help with sales.” Prefer prompts like “Review new inbound companies, check whether the account matches ICP criteria, summarize buyer context, and prepare a concise owner handoff. Ask for approval before changing deal stage or sending external messages.”

Review quality

Review the run summary, status, tool calls, credit usage, and output location. If the agent failed, check whether the failure came from missing context, missing permissions, unclear instructions, or an unavailable integration.

Related pages

Last updated