Research Agent
You need a researched answer to an open question. Your mission: an agent that plans its own searches, reads what it finds, and writes a brief with checkable citations — under a budget and bounds. Not a script looping over Google — a system that makes decisions, knows when to stop, and admits when it doesn't know.
What you'll build
A full agentic loop written by hand: planning, search & read tools, a complete ReAct trace, iteration/budget bounds with a give-up exit, and a synthesized brief with verified citations — with reliability and cost engineering.
Why it matters
The pre-capstone summit: the full agentic loop, the single most-asked-for skill in AI-engineering interviews. People who build it with a framework without understanding the loop itself fail the first 'so what do you do when it won't stop?' question. You'll have written it by hand — so you'll know.
Skills you'll gain
Prerequisites
PRJ-3 evaluated + gate M7 + a web-search key.
Requirements
- RQ-1: the agent stops at a declared iteration/budget cap — always, not usually.
- RQ-2: every claim in the brief has a checkable source — a real URL that actually contains the cited content.
- RQ-3: cost per run is measured and stays under the cap.
- RQ-4: every run has a complete ReAct trace — each thought, action, and observation logged and reviewable.
- RQ-5: a give-up exit: if the question can't be answered within bounds, the agent says so explicitly instead of fabricating or spinning.
- RQ-6: the core loop is hand-written — no agent framework inside it.
Architecture
Product-level: question → planner → the loop (thought → action → observation) → search & read tools → synthesizer → cited brief — all inside a frame of iteration meter, cost meter, and trace log. The explanation teaches why the loop is an engineered object with bounds and stop conditions, not which framework to use.