On this page
Learning objectives
- Specify and build a bounded evidence-gathering agent
- Require citations and human approval before a brief is released
- Evaluate quality, cost, and recovery with realistic cases
Prerequisites
Modules 00–10 · An approved source list · A real research brief template
Answer first
Canonical definition
The capstone agent turns a defined marketing question into a cited draft research brief by searching approved sources, extracting evidence, identifying gaps, and stopping for analyst review before any downstream action.
Operating context
Why it matters
This project joins the Agent Building and AI Marketing tracks around a real workflow rather than a toy chat or weather demo.
The result is valuable even with limited autonomy because traceable evidence and faster analyst review solve an operating problem.
Method
Step by step
- 01
Write the agent spec
Define question, source policy, tools, evidence schema, step budget, completion criteria, prohibited actions, and reviewer role.
- 02
Build the evidence loop
Search approved sources, extract claim-level evidence, deduplicate, score gaps, and preserve citations in application state.
- 03
Draft and review
Generate a brief that separates facts, customer language, inference, uncertainty, and unanswered questions, then route it to an analyst.
- 04
Evaluate and operate
Test source quality, citation fidelity, coverage, permission boundaries, cost, repeated actions, and recovery from tool failure.
Hands-on lab
Ship the complete capstone
Implement the agent in a staging environment against one real, non-sensitive research question.
Deliverable
Repository, architecture note, agent spec, eval set, run traces, analyst review UI, cost report, and production readiness decision.
What breaks
Common failure modes
- F1Allowing unrestricted web sources without a documented evidence policy
- F2Producing a polished report that loses claim-level citations
- F3Adding publishing or outreach authority before research quality is proven
Beyond the demo
Production notes
- Start read-only and analyst-facing. A useful research copilot is a valid outcome even if no action is automated.
- Use real rejection reasons to improve the eval set before changing prompts or increasing autonomy.
Further reading
Sources
Related Tenten resources