fulcra-agent-atc

# Fulcra Agent ATC A subscription fleet — Claude Max, OpenAI/Codex plans, maybe a local ollama and a hosted overflow lane, across one or several harnesses — has two failure modes that waste money and stall work: it **burns frontier budget on mechanical work** a cheap model would finish, and it **stalls on one capped account** — hitting a window cap on one subscription while another sits idle. ATC (air-traffic control) is the routing layer that fixes both, and it runs on **one account** the same way it runs on a fleet. Three parts: - **The map** — a packaged default model map (`map_version 2026-07-08`) tagging every model by capability (`code`, `architecture`, `writing`, `long-context`, `vision`, `fast`, `tool-use`) and a cost rank (1 = scarcest cap-weight, 9 = locally free). You declare accounts; the map says which models each can run and what they cost. - **The router** — `coord-engine route <team> --needs …`: ranks the models that cover your declared needs **cheapest-capable-first**, filtered to the accounts that have headroom right now. `coord-engine headroom` is the shared cap ledger it reads. - **The feedback loop** — `coord-engine usage log … --model --task-class --outcome`: every dispatch logs what it cost and how it turned out. Repeated bad outcomes on a (model, task_class) pair **demote** it — the router marks it and ranks it below everything that hasn't been failing. Dispatch itself stays native to your harness — ATC decides *which model against ...