Instructions

Instructions define how an agent should behave. Use them for durable policy, role, tone, and workflow rules that apply to every run.

Add stable instructions

const agent = new AgentBuilder('support', model)
  .instructions([
    'Answer from verified support information.',
    'Use tools before making account-specific claims.',
    'Escalate billing, legal, or security uncertainty.',
  ].join('\n'))
  .build()

Good instructions describe repeatable behavior. They should not contain the current user's permissions, tenant ID, database IDs, or other request-specific facts.

Compose instruction blocks

Multiple .instructions(...) calls append blocks in order; they do not replace earlier instructions.

const agent = new AgentBuilder('research', model)
  .instructions(baseAgentRules)
  .instructions(researchWorkflowRules)
  .skills(researchSkills)
  .build()

Anvia combines normal instruction blocks first, followed by skill instructions. Keep the order intentional and remove contradictory rules.

Choose the right boundary

Information	Put it in
Durable behavior and policy	Instructions
Facts the model should use	Context
User or tenant identity	Session and runner state
Product permissions	Tool handlers and services
Trace correlation	.withTrace(...)

If one request needs fundamentally different behavior, create a different agent or scoped factory instead of appending conflicting instructions at runtime.

Keep enforcement in code

Instructions can tell a model when to use a tool, but they cannot enforce safety. Every side-effect tool must still validate the user, tenant, input, permissions, and approval state.