Skip to main content
The instructions() helper creates well-structured, consistent prompts for agents. It formats instructions into a standardized template with purpose and routine sections.

​
Function Signature

agent.ts:377:395
export function instructions({ purpose, routine }: PurposeRoutineInstructions): string

​
Parameters

​
purpose
string | string[]
required
The agent’s role or purpose. Can be a single string or array of strings.
// Single string
purpose: 'You are a helpful assistant'

// Multiple strings
purpose: [
  'You are a data analyst.',
  'You specialize in financial data.',
]
​
routine
string[]
required
Step-by-step workflow for the agent. Array of strings.
routine: [
  'Analyze the user request',
  'Gather necessary data',
  'Provide a clear response',
]

​
Return Value

Returns a formatted string: 
# Agent Context
[purpose lines]


<specialized_agents_placeholder>

Use the following routine to fulfill the task.
# Routine
1. [first routine step]
2. [second routine step]
...
The <specialized_agents_placeholder> is replaced with handoff agent information when the agent has handoffs configured.

​
Variants

​
instructions.swarm()

For multi-agent coordinators. Includes additional prompting about the multi-agent system:
agent.ts:397:418
instructions.swarm({ purpose, routine })
Adds a system context prefix: 
# System context
You are part of a multi-agent system called the DeepAgents SDK...
Handoffs are achieved by calling a handoff function, generally named `transfer_to_<agent_name>`.
...

# Agent Context
[purpose]

<specialized_agents_placeholder>

# Routine
...

​
instructions.supervisor()

For supervisor agents that coordinate multiple specialists:
agent.ts:420:444
instructions.supervisor({ purpose, routine })
Adds supervisor-specific directives: 
# System Context
You are part of a multi-agent system called the DeepAgents SDK...

Core Directives:
- Begin with a concise checklist (3-7 bullets) of what you will do
- Continue working until the user's query is completely resolved
- Your thinking must be thorough and step-by-step
...

# Agent Context
[purpose]

<specialized_agents_placeholder>

# Routine
...

​
instructions.supervisor_subagent()

For agents that work under a supervisor:
agent.ts:446:478
instructions.supervisor_subagent({ purpose, routine })
Adds automatic transfer back to supervisor at the end of the routine.

​
Examples

​
Basic Usage

import { agent, instructions } from '@deepagents/agent';
import { openai } from '@ai-sdk/openai';

const dataAnalyst = agent({
  name: 'data_analyst',
  model: openai('gpt-4o'),
  prompt: instructions({
    purpose: [
      'You are a data analyst specializing in business intelligence.',
      'You help users understand their data through clear insights.',
    ],
    routine: [
      'Examine the provided data carefully',
      'Identify key patterns and trends',
      'Calculate relevant statistics',
      'Present findings in a clear, actionable format',
    ],
  }),
});

​
Multi-Agent Coordinator

import { agent, instructions } from '@deepagents/agent';

const researcher = agent({
  name: 'researcher',
  model: openai('gpt-4o'),
  prompt: 'Research topics thoroughly.',
  handoffDescription: 'Handles research tasks',
});

const writer = agent({
  name: 'writer',
  model: openai('gpt-4o'),
  prompt: 'Write engaging content.',
  handoffDescription: 'Handles writing tasks',
});

const coordinator = agent({
  name: 'coordinator',
  model: openai('gpt-4o'),
  prompt: instructions.swarm({
    purpose: ['Coordinate research and writing to complete user requests'],
    routine: [
      'Analyze the user request',
      'Use transfer_to_researcher for information gathering',
      'Use transfer_to_writer for content creation',
      'Review and refine the output',
    ],
  }),
  handoffs: [researcher, writer],
});

​
Supervisor Pattern

import { agent, instructions } from '@deepagents/agent';

const supervisor = agent({
  name: 'supervisor',
  model: openai('gpt-4o'),
  prompt: instructions.supervisor({
    purpose: [
      'You are a project supervisor coordinating multiple specialist agents.',
      'Delegate tasks and review results to ensure quality.',
    ],
    routine: [
      'Break down the user request into subtasks',
      'Delegate each subtask to the appropriate specialist',
      'Review specialist outputs',
      'Synthesize results into a cohesive response',
    ],
  }),
  handoffs: [specialist1, specialist2, specialist3],
});

​
Research Bot Example

From the source code:
research_bot.ts:48:53
const planner = agent({
  model: openai('gpt-4o'),
  name: 'PlannerAgent',
  output: WebSearchPlanSchema,
  prompt: instructions({
    purpose: [
      'You are a helpful research assistant. Given a query, come up with a set of web searches to perform to best answer the query.',
    ],
    routine: ['Output between 5 and 10 terms to query for.'],
  }),
});

​
Financial Analysis Example

From the source code:
finanicals_bot.ts:73:83
const riskAgent = agent({
  name: 'RiskAnalystAgent',
  model: groq('openai/gpt-oss-20b'),
  output: AnalysisSummarySchema,
  prompt: instructions({
    purpose: [
      "You are a risk analyst looking for potential red flags in a company's outlook.",
      'Given background research, produce a short analysis of risks such as competitive threats, regulatory issues, supply chain problems, or slowing growth.',
    ],
    routine: ['Keep it under 2 paragraphs.'],
  }),
});

​
Handoff Agent List

When an agent has handoffs, the <specialized_agents_placeholder> is replaced with a formatted table: 
## Specialized Agents

| Agent Name | Agent Description |
| --- | --- |
| researcher | Handles research and fact-finding tasks |
| writer | Handles writing and content creation |
This is automatically generated from the handoffDescription of each handoff agent.

​
Custom Prompt Formats

You can combine instructions() with additional custom text: 
const customAgent = agent({
  name: 'custom_agent',
  model: openai('gpt-4o'),
  prompt: [
    instructions({
      purpose: ['You are a helpful assistant'],
      routine: ['Help the user', 'Be clear and concise'],
    }),
    '',
    '## Additional Guidelines',
    '- Always be polite',
    '- Verify facts before stating them',
    '- Ask for clarification when needed',
  ].join('\n'),
});

​
Dynamic Instructions with Context

Combine with context-aware prompts: 
interface UserContext {
  userId: string;
  expertiseLevel: 'beginner' | 'intermediate' | 'expert';
}

const tutorAgent = agent<unknown, UserContext>({
  name: 'tutor',
  model: openai('gpt-4o'),
  prompt: (ctx) => {
    const baseInstructions = instructions({
      purpose: ['You are a programming tutor'],
      routine: [
        'Assess the user\'s question',
        'Provide clear explanations',
        'Give relevant examples',
      ],
    });
    
    const levelGuidance = ctx?.expertiseLevel === 'beginner'
      ? '\n\nUse simple language and avoid jargon.'
      : ctx?.expertiseLevel === 'expert'
      ? '\n\nYou can use technical terminology and advanced concepts.'
      : '\n\nBalance simplicity with technical accuracy.';
    
    return baseInstructions + levelGuidance;
  },
});

​
Best Practices

Clearly define what the agent does:
// ✅ Good - specific
purpose: [
  'You are a financial analyst specializing in tech companies.',
  'You analyze earnings reports and market trends.',
]

// ❌ Vague
purpose: 'You help with finance'
Make routine steps concrete and actionable:
// ✅ Good - actionable
routine: [
  'Read the user\'s question carefully',
  'Search for relevant documentation',
  'Provide code examples from the docs',
  'Verify the examples work',
]

// ❌ Vague
routine: ['Help the user', 'Do a good job']
Choose the right variant for your use case:
  • instructions() - Single agent
  • instructions.swarm() - Multi-agent coordinator
  • instructions.supervisor() - Supervisor coordinating specialists
  • instructions.supervisor_subagent() - Specialist under a supervisor

​
See Also

agent()

Create agents

Multi-Agent Guide

Build multi-agent systems

Build docs developers (and LLMs) love

Get started for freeTalk to us