Skip to main content
Custom agents are purpose-built with a defined role, goal, and task set. Each agent runs against the same knowledge graph and tool library as the built-in agents, scoped to a specific, repeatable workflow.

​
Agent types

TypeAccess
PrivateAccessible only to the creator. Use for personal workflows and agents under active development.
SharedAccessible to specific team members granted access by email. Collaborators can execute the agent but cannot modify its configuration.
PublicAccessible to all members of the organization. Discoverable in the agent library for standardized, reusable workflows.

​
Agent definition

Every custom agent specifies four attributes:
  • Role: The professional function the agent adopts, such as "Senior Python Developer" or "Security Auditor". This shapes how the agent frames its analysis.
  • Goal: The primary objective, defined with specific, measurable outcomes, such as "Identify SQL injection vulnerabilities across all database query constructions".
  • Backstory: Professional context that informs decision-making. Include relevant methodology, standards, or principles the agent applies.
  • System Prompt: High-level instructions that control execution behavior: output format, edge case handling, and quality standards.

​
Task configuration

Each agent contains between one and five tasks. A task defines a discrete action and consists of three components:
  1. Description: A clear statement of what the task accomplishes, focused on outcomes. The system automatically enhances descriptions with step-by-step reasoning to improve execution quality.
  2. Tools: An array of tool IDs the agent can use. Each task requires at least one tool, selected from the library of 70+ options.
  3. Expected Output: A JSON object specifying the format and structure of the result.

​
Execution

Custom Agents List Potpie runs each task in sequence using the configured tools and streams the result with file and line citations. Complex queries requiring synthesis across multiple files trigger multiagent mode automatically. A supervisor agent coordinates the work and delegates to specialized subagents.

​
Agent generation

Generate a custom agent from a natural language description. The system infers an appropriate role, goal, backstory, and task breakdown, and recommends tools matched to the described workflow. The agent executes immediately and accepts refinements through subsequent attribute updates.