Energy & Environment Lab
EcoScout: AI Agents for Government Contact Data
With Vishal Vincent Joseph, Data Science Research Manager at the E&E Lab
Multi-agent pipeline that mapped 5,501 local government bodies across nine African countries, collecting contact details, websites and social media for mayors, governors, council chairs, environmental regulators and planning authorities.
Frame country. A person decides what the country’s local governments even are: tiers, units, executives and what each body is called locally.
Context
A potential intervention would deliver air-quality information directly to local government offices and measure its effect on outcomes. That starts with knowing who those offices are and how to reach them, across countries whose governments are structured and named in very different ways.
Approach
- Framed each country by hand first: a governance profile of its tiers, units, executives and environmental roles, checked against legal and statistical sources
- A discovery agent searches the web for candidate websites for each institution, and a judge decides which one is correct, if any
- Deterministic scraping code reads the sites that were found, and a separate contact agent looks up whatever is still missing
- Code verifies every result at the end: domain lookups, phone checks and HTTP status checks on every site
- Kept each agent to a narrow job and did as much as possible in plain code, since an agent given too many tasks fails on at least one
Outcome
- 5,501 bodies and 5,331 executive seats mapped across nine countries, using 1,021 agent runs and 10,662 web searches in about 26 hours of wall-clock time
- The channel that works is country-specific: South Africa is 98% reachable by phone, Angola 91% by social media, Tanzania 86% by website
- In DR Congo, 1,054 of 1,061 government bodies have no discoverable web presence, each one searched and confirmed