AI Agents Foundations: Build a Bounded Research Agent

The project

Build an agent that answers a research question from a fixed list of approved documents. It may search document titles and return citations. It must not browse the web, send messages, change files, or infer facts that the tool did not return.

Step 1: write the contract

Define input, permitted tools, state, output, and stop conditions. A useful first contract is: question in, up to three document lookups, answer with quoted document identifiers, then stop. State should record the question, tool calls, observations, and final answer separately.

Step 2: implement one read-only tool

Use a function such as search_documents(query) that returns a small typed result. Reject empty queries and cap results. The agent receives tool output, not direct database access. This keeps both testing and later permission review manageable.

Step 3: add a human checkpoint

Before any future write action, present the proposed action, affected target, and reason. For this first project, the checkpoint always stops the agent after its answer. A human can inspect the trace and decide whether the contract was followed.

Practice: evaluate five cases

Create five questions: two answerable, one ambiguous, one with no supporting document, and one instruction-injection attempt. Record whether the agent cited its evidence, stopped within its tool budget, and refused unsupported claims.

Common mistake

Do not call a multi-step loop an agent merely because it has a prompt. The useful distinction is controlled state plus permitted actions plus evidence you can inspect.