Prompt file imported from doubts-suplab/eeik-bootstrap (.github/prompts/workflows/multi-agent-system-design.prompt.md). Copyright stays with the author.

Multi-Agent System Design Workflow

Design a multi-agent AI system to solve the following problem.

Problem: $PROBLEM_DESCRIPTION Framework Preference: LangGraph / CrewAI / AutoGen / Custom A2A

Workflow Steps

Step 1 — Problem Decomposition

Before designing agents, decompose the problem:

• What are the distinct subtasks that must be completed?
• Which subtasks can run in parallel vs. must be sequential?
• What information must flow between subtasks?
• Where do human decisions need to be inserted?

Step 2 — Agent Topology Design

Define the agent architecture:

• List each agent with: name, role, capability scope, tools available
• Define the orchestration pattern (supervisor, peer-to-peer, pipeline)
• Design the communication protocol (task schema, context object)
• Identify where output validation between agents is needed

Step 3 — State & Context Design

Design the shared state:

• Define the typed state schema (Pydantic model / TypedDict)
• Identify which state fields accumulate (require reducers)
• Design context pruning for long chains (token budget management)
• Define checkpoint strategy for resumability

Step 4 — Safety Design

Before implementation, define:

• Maximum iteration count for any agent loop
• Timeout per agent call
• Human-in-the-loop gates and their trigger conditions
• Rollback/abort mechanism for stuck pipelines
• Output validation before passing to downstream agents

Step 5 — Implementation

Activate the appropriate framework agent:

• LangGraph → langraph-engineer agent
• CrewAI → crewai-engineer agent
• AutoGen → autogen-engineer agent
• Custom A2A → a2a-engineer agent

Step 6 — Observability Setup

• Configure LangSmith / logging for all agent interactions
• Implement token usage tracking across all agents
• Produce a test harness with representative inputs

Output

Agent topology diagram, typed state schema, complete implementation, and safety controls documentation.