Concepts & Vocabulary

June 25, 2026 · View on GitHub

Fleet engineering sits one layer above loop engineering. This glossary links the stack so you can design fleets with the right mental model.

Fleet Engineering

Replacing ad-hoc agent populations with accountable organization. You design registry, identity, permissions, inbox, audit, economics, and sovereign control — so many loops and agents can run across a team without chaos.

A fleet is a governed population, not a folder of prompts.

The Accountability Test

If you cannot complete this sentence for any action, you have a population — not a fleet:

Which agent did it, with what authority, against what task, evidenced by what?

See accountability-test.md.

Context Engineering

What the model sees at each inference. Subset of harness; prerequisite for any agent.

Harness Engineering

The environment one agent run executes in: tools, sandboxes, memory, verification, permissions.

Harness = single session / single run
Loop    = harness + schedule + state + verification chain
Fleet   = many loops/agents + registry + identity + org-wide control

Loop Engineering

The system that replaces you as the prompter for one autonomous workflow. See loop-engineering.

Fleet engineering begins when:

Multiple loops collide or compete for resources
Multiple people create agents without a catalog
Compliance asks "who did this?"
Token spend is nobody's job to watch

Population vs Fleet

Population	Fleet
Agents exist because someone made them	Agents are registered with owners
Credentials are shared or unknown	Identity model is explicit (claw vs assistant)
Permissions are implicit	clone / run / edit is enforced
Debugging is per-agent	Audit trail is cross-agent
Kill = delete a chat	Kill switch affects subsets or whole fleet

Claw vs Assistant (LangSmith Fleet terms)

Claw: fixed credentials; acts as a team resource (e.g. shared Linear bot)
Assistant: acts on behalf of invoking user; OAuth per user; permissions follow the human

These map to identity & credentials — one of the five fleet concerns.

Runnable companion: assistant-vs-claw — mock Notion/Slack demos, trust-boundary tests, and a Fleet mapping guide. No API keys required.

The Five Concerns (+ Registry)

See five-concerns.md.

Topology — how agents relate (hierarchical, P2P, blackboard, router)
Choreography — how work flows (workflows, events, handoffs)
Identity & trust — who acted, on whose behalf
Resource economics — budgets, quotas, admission control
Sovereign control — kill switches, rollback, autonomy tiers

Registry is the spine that makes the five concerns operable at scale.

Concept Map

flowchart TB
    CE[Context Engineering] --> HE[Harness Engineering]
    HE --> LE[Loop Engineering]
    LE --> FE[Fleet Engineering]
    FE --> R[Registry]
    FE --> I[Identity]
    FE --> P[Permissions]
    FE --> IN[Inbox]
    FE --> AU[Audit]
    FE --> EC[Economics]
    FE --> SC[Sovereign Control]

Maturity Levels (F0–F3)

Level	Name	Posture
F0	Ad-hoc	Agents everywhere; no catalog
F1	Cataloged	Registry + permissions doc; inbox for risky actions
F2	Shared	Team agents, budgets, cross-agent audit
F3	Enterprise	Policy-as-code, kill switches, SLOs, compliance

See maturity-model.md.

Loop ↔ Fleet Maturity Bridge

Fleet engineering assumes at least one stable loop. Do not run unattended L2+ loops on an F0–F1 fleet.

Loop level	Fleet level	Safe together?	Notes
L0 manual	F0 ad-hoc	Yes	Default chaos — acceptable for experiments
L1 scheduled	F1 catalog	Yes	Registry + human gates on risky actions
L2 unattended	F1	No	Need F2 inbox + budget before unattended writes
L2+ unattended	F2	Yes	Per-agent caps, audit path, kill switch documented
L3 enterprise	F3	Yes	Policy-as-code on both layers

Promotion rule: raise fleet tier before loop tier when loops touch production or external systems.

Starter: starters/fleet-plus-loop.