One always-on system of specialist agents — orchestrated by a Chief, challenged by an independent Investment Committee, and grounded in a single living knowledge base — supporting the development of a permitted LNG plant through to financing, EPC award and FID.
Every email, permit, model and call note lands in a raw archive that is never edited. An LLM compiles it into an interlinked wiki that every agent reads from, with each fact tagged verified, estimate, assumption or open question — so guesses never compound into "facts."
A Chief Development Agent routes and reconciles work across Intelligence, Finance, Capital Markets, Commercial, Engineering, Legal, Regulatory, EPC, Risk and Document Control — while a separate Investment Committee deliberately looks for reasons the deal should fail at every gate.
Scheduled runs track regulatory dockets, cost of money, LNG pricing, the competitive landscape, EPC and supply chain, and counterparties — updating the brain and surfacing only what crosses a real threshold. A separate self-improvement loop refines each agent's own playbook, gated by human approval.
The team works toward a sequence of real decisions — project baseline and mandate, development capital raise, commercial and technical bankability, then financing, EPC award and FID — with the Investment Committee issuing pass, conditional pass or hold at each one.
Runs on
Claude Code or a small local web app; one daily brief, to the project owner only.
Claude Agent SDK — the Chief as orchestrator, each specialist a subagent defined by a charter and skills.
A single always-on Mac mini, with encrypted offsite backup and remote access over Tailscale.
Team structure, the living brain, watch loops, gates and the Mac mini build — in detail.