How to use the AI Agent Architecture Planner
- 1. Describe the jobSay what the system should do and answer questions about steps, actions, data, speed and risk.
- 2. Get a planSee the recommended pattern, why, a flow diagram, components and guardrails.
- 3. Download the docTake a Markdown design doc to your team and start with a small prototype.
The patterns
Not every AI feature needs an agent. The planner chooses between six common patterns. A single LLM call suits simple tasks. A prompt chain runs fixed steps in order, and a router sends each request to a specialised path. RAG answers from your documents. A single agent with tools lets the model choose actions in a loop, and an orchestrator with workers splits big jobs into subtasks. The rules follow the workflow and agent patterns in Anthropic's widely cited Building effective agents guide: use fixed workflows when the steps are predictable and agents only when they aren't.
Worked example
The default describes a course assistant that answers student questions and books counselling calls. Its steps vary, it changes things (bookings), it uses documents and calls other APIs, it must reply in real time, mistakes cost a medium amount and it handles personal data. The planner recommends a single agent with tools, adds retrieval for the documents and lists 6 guardrails, including approval before write actions and handling personal data under India's DPDP Act. Change it to fixed steps, no actions and no APIs, and it recommends RAG instead.
Before you build
- Collect 20–50 real example tasks with expected results; they become your test set.
- Build the simplest version end to end, measure it, and add complexity only where it helps.
- Log every model and tool call; agents are hard to debug without traces.
- Treat documents and web pages as untrusted input that may contain prompt injection.
Frequently asked questions
Is the recommendation always right?
No. It's a starting point from rules we wrote. Prototype and measure before you commit to an architecture.
Which framework should I use?
Any can implement these patterns: plain API calls, LangChain/LangGraph, or agent SDKs from model providers. Our LangChain code generator gives starter code.
Is the AI agent architecture planner free?
Yes. You can open the AI agent architecture planner and fill in everything without signing up. When you get a plan for the first time, we ask once for your name and WhatsApp number; after that it, and every other Uncodemy tool, opens straight away.