Orchestrating Agents:
The System Prompt as a Manager

An AI agent is a prompt with tools. A multi-agent system is a team of prompts with a manager. The manager prompt is the most important piece.

The Rise of Agentic AI

We've moved from single-turn chatbots to autonomous agents that can:

• Browse the web and gather research
• Write and execute code
• Coordinate with other agents to solve complex tasks
• Make decisions and take actions over multiple steps

But here's the secret: the agent's behavior is defined entirely by its system prompt. A poorly written system prompt creates a chaotic, unreliable agent. A well-written one creates a focused, effective team member.

Anatomy of a Manager Prompt

The "Manager" agent coordinates other "Worker" agents. Here's the blueprint:

# System Prompt: Research Manager Agent

## Role
You are a Research Manager coordinating a team of specialist agents.

## Your Team
- **Web Researcher**: Searches the internet for current information
- **Data Analyst**: Processes numbers, creates charts
- **Writer**: Produces final reports in markdown

## Workflow
1. Break the user's request into subtasks
2. Assign each subtask to the appropriate specialist
3. Review each specialist's output for accuracy
4. Synthesize results into a final deliverable

## Constraints
- Never fabricate data. Always cite sources.
- If a specialist returns low-confidence results, ask them to retry.
- Maximum 3 delegation rounds before delivering best-effort result.

## Handoff Format
When delegating, use this format:
{
    "agent": "web_researcher",
    "task": "Find 2024 revenue for Company X",
    "context": "For a competitive analysis report",
    "deadline": "Return within 2 turns"
}

The 5 Rules of Agent Prompts

Rule 1: Define the Role Precisely

"You are a helpful assistant" is not a role. "You are a senior financial analyst who specializes in SaaS metrics" is.

Rule 2: Enumerate Available Tools

Agents need to know exactly what they can and cannot do:

const agentPrompt = await promptOps.getPrompt("code-reviewer", {
    environment: "production"
});

// The prompt defines available tools:
// ✅ read_file, search_codebase, create_comment
// ❌ delete_file, push_to_main, modify_permissions

Rule 3: Set Explicit Constraints

Without boundaries, agents will surprise you:

• Budget limits: "Do not make more than 10 API calls per task."
• Scope limits: "Only operate on files in the /src directory."
• Safety limits: "Never execute DELETE operations without human approval."

Rule 4: Define the Handoff Protocol

When agents talk to each other, they need a structured communication format. JSON schemas, not free-form text.

Rule 5: Version Everything

If your manager prompt changes, every agent's behavior changes. This is why prompt versioning is critical for multi-agent systems. One untested prompt update to the manager can cascade failures through the entire agent team.

Why Versioning Matters More for Agents