Project: Multi-Step AI Agent
A multi-step agent is a loop, not a magic prompt.
text
goal -> plan -> tool/action -> observe -> update -> verify -> stop
Project goal
Build an agent that can research a topic and produce a short sourced brief.
Required pieces
| Piece | Purpose |
|---|---|
| planner | creates steps |
| tool layer | search or retrieve information |
| memory/state | tracks what is done |
| verifier | checks sources and claims |
| final writer | summarizes with citations |
Safety boundaries
For a first agent:
- read-only tools only
- max 5 tool calls
- no external messages
- cite sources
- stop if evidence is missing
Success criteria
The agent succeeds when it returns:
- answer
- key evidence
- citations
- uncertainty
- next questions
Knowledge check
Q1: Why does an agent need state?
To track progress and avoid repeating work.
Q2: What should the agent do when evidence is missing?
Stop or ask for more information, not hallucinate.