Skip to main content

​
Custom agents

Custom agents turn reusable instructions into a first-class Starfire AI object. Instead of recreating the same role and operating rules in every conversation, an agent can carry a durable purpose, behavior, and capability boundary.

​
What an agent represents

An agent can define concepts such as:
  • name and description
  • durable instructions
  • intended use case
  • preferred model or model eligibility
  • tool access
  • knowledge or project context
  • limits and budgets
  • organization ownership
  • runtime permissions
The exact fields available depend on the current Alpha build.

​
Agent vs chat

A chat is a conversation instance. An agent is a reusable configuration that can start or participate in conversations and workflows.

​
Good agent boundaries

Agents are strongest when they have a coherent job. Examples:
  • CodeForge — software-development and debugging work across supported languages and frameworks.
  • release manager — turns completed engineering work into release plans and validation checklists.
  • documentation agent — drafts product documentation using a defined terminology and structure.
  • support triage agent — organizes a support report and gathers diagnostics without receiving unrestricted administrator access.
One giant “do everything” agent is harder to reason about and harder to permission safely.

​
Tools and permissions

An agent should receive only the tools needed for its role. A documentation agent may need Knowledge and web search but not build filesystem access. A FORGE-oriented agent may need project files and validation tools. Tool access should be bounded by the active user, organization, plan, and administrator configuration. An agent does not become more privileged than its execution context simply because its instructions ask for more access.

​
Knowledge and projects

Agents can become more useful when attached to durable project or Knowledge context. That gives the agent relevant reference material without embedding large instructions into the agent definition itself.

​
From agents to automation

NEXUS expands the concept beyond reusable chat personalities. Agents can become participants in tasks, workflows, schedules, watchers, and approval-controlled automation as those surfaces are enabled.
Custom agents are an active Starfire product capability. Deeper autonomous task/workflow execution is part of the NEXUS-era Alpha expansion and can be staged separately.

Writing instructions

Design clear, durable behavior without overloading the agent prompt.

Tasks & workflows

Understand how agents participate in multi-step work.

Schedules & watchers

Learn the event and time-based automation model.

Approvals & memory

See how human approval and persistent memory constrain automation.