Back to All Case Studies
energyFull Access Unlocked

Eight Specialist Agents That Screen a Data-Centre Site in One Run

Pre-due-diligence on a candidate site used to mean weeks of separate specialist reviews in incompatible formats. We built a swarm of eight domain agents that queries real public data and assembles one scored, sourced report.

European energy major · data-centre development

energy
8
specialist Agents
19
report Sections
4
weeks To V1

!The Need

Screening a candidate co-location site before due diligence means assessing land, transport, grid, environment, water, planning and fibre — each by a different specialist, in a different format, on a different timeline. Sites could not be compared on a like-for-like basis, so early ranking came down to judgement. Worse, community acceptance and ESG — the dimension most likely to kill a project late and expensively — was assessed last, or not at all, because it was the hardest to quantify.

The Approach

We built a swarm of eight specialist agents on the Claude Agent SDK, one per diligence domain, with community and ESG promoted to a first-class agent rather than an afterthought. Each agent queries real public sources — cadastral registries, open infrastructure and grid maps, OpenStreetMap, the European environment agency's protected-area service, internet-exchange peering data and crawler-wrapped government portals — instead of relying on model recall. Every agent output, its evidence and its score persist to their own rows, and a narrative assembler composes them into a single structured report. Eight golden-site fixtures run as graded evaluations in CI, so a prompt change that quietly degrades report quality fails the build rather than reaching a decision-maker.

Technologies Used

Claude Agent SDKTypeScriptSupabaseReactVitestGitHub ActionsFirecrawl

The Output

Eight specialist agents — land, transport, grid, environment, water, planning, fibre, and community/ESG
A 19-section pre-due-diligence report per run, with a composite score and per-agent scores and observations
Over 40 ranked next-step actions carried into the following diligence stage
Every agent output, evidence item and score persisted and auditable after the fact
Eight golden-site fixtures graded on every CI run, so quality regressions fail the build
An auth-gated report viewer rendering live engine output, not static slides

The Impact

Version one delivered in four weeks, from empty repository to reports a decision-maker reads
On the benchmark site the engine scored 63.05 against an independent engineering firm's reference figure of 62 — roughly one point apart, with no analyst in the loop
Scores discriminate rather than cluster: across eight reference sites results span 47.96 for a rural location to 75.5 for a prime corridor
Community and ESG risk is now scored on the first pass, at 63.6% differentiator coverage, instead of surfacing after money is committed
Site comparison moved from incompatible specialist documents to one consistent, sourced score

Ready to achieve similar results?

Contact Us
Supalabs AI solutions