Fleet

Overview

Fleet is a family of starter skills tiered by task complexity. Instead of one monolithic prompt, a fleet worker loads the tier that matches the task — L1 through L5 — so the injected context budget scales with the work: a single-step command gets a minimal seed; a complex self-managed task gets a compact prompt with KB pre-fetch. Each level tells the worker how much to decompose, whether to retrieve KB context, and when to escalate a level-mismatch.

The five levels

• L1 — Basic — a single-step command with no decision-making. One tool call, no RAG, no KB lookup. Minimal injected context (~200 tokens). If the task turns out to need analysis or multiple steps, stop and report level-mismatch: needs >=L3 instead of improvising.
• L2 — Structured — a few templated steps. Stay within the role's allowed-command list; no open-ended analysis. No RAG unless the brief explicitly references a KB document. Return a short structured summary.
• L3 — Analyst — analysis with variability and a choice between alternatives. Retrieve KB context on demand (semantic lookup) only for the specific points needed — do not preload. State the chosen option and a one-line rationale.
• L4 — Expert — a complex task with several subtasks. Decompose before acting, keep a running subtask checklist, retrieve KB context on demand per subtask. Return a structured summary: subtasks done, outcomes, open items.
• L5 — Autonomous — a complex, self-managed task; the worker coordinates its own subtasks and may sequence multiple agents' worth of work. Pre-fetch the KB anchors for the domain, then work autonomously within role permissions — never taking a hard-gated action (prod-deploy, secret-rotation) without operator approval.

Why tiered context

Matching the injected context to the task level keeps the token budget proportional: elementary work does not pay for retrieval it never uses, and complex work is not starved of the decomposition guidance and KB pre-fetch it needs. The level-mismatch escalation keeps a worker from improvising beyond its tier — an L1 worker that hits real analysis stops and asks for a higher tier rather than guessing.