Entity Screening Agent
Agentic web research for entity screening: structured outputs, exhaustive search patterns, and a trace of each step.
- Role
- Lead software & AI
- Organization
- Financial services
- Context
- Professional
- Capabilities
- Agent orchestration, Web research, Entity resolution, Observability, Evaluation
Overview
Entity screening fails in two opposite ways: a miss, or a pile of unstructured pages nobody will read. This work treated screening as an agent loop with a trace — every search, every extraction, every structured field — rather than a single prompt.
I implemented the agentic workflow and the trace/output shape inside a consulting team. Watchlists, vendor contracts and compliance sign-off were not mine.
Problem
Analysts needed exhaustive-enough web research on organisations, with results that could be handed on as structured records instead of chat logs.
Approach
A bounded agent: plan searches, fetch, extract into a schema, log each step, stop. The interesting part is the trace, not the model. Exact dates, vendors and hit rates stay off this stealth page.
Tradeoffs and limitations
Exhaustiveness is expensive and never complete. Traces help a reviewer and also retain more of the web than a summary would. I do not claim coverage percentages or time saved.
A later private codebase continues this line of work. It is not published here.
Confidentiality
Stealth / capability-only. No client name, no screenshots, no unique metrics.