Environment provisioning agent
Developers wait days for new environments, so you're building a provisioning agent: ask it for an environment, get one. This thing touches your cloud account, so be precise about the permission model — what it can create, with what limits, and what it must never be able to do even if prompted maliciously. Include cost controls. How would you test the safety boundary as rigorously as the happy path, and what's the pitch to the platform team?
screen_provision_request(request_fields: list[str], team_active_envs: int) → str[["size=medium","region=us-east","ttl_hours=24","budget_usd=200","network=dev"],2]out"approve"[["size=small","region=eu-west","ttl_hours=8","budget_usd=50","network=prod"],0]out"deny:network"[["size=large","region=us-east","ttl_hours=48","budget_usd=900","network=staging"],1]out"escalate:budget_usd"Turn the fuzzy goal into a bounded agent task. Say what the agent reads, what it may do on its own versus draft for a human, and what it must never touch. Then earn the trust: how you’d verify it before rollout, and the pitch (the problem, what you built, why it matters).
Agent build: design the agent in plain language (or narrate it). The coach grades the decomposition, the guardrails, the verification plan, and the pitch.