A trail from prompt to fleet. Each layer owns a different question.
flowchart LR
P[Prompt Engineering] --> C[Context Engineering]
C --> H[Harness Engineering]
H --> L[Loop Engineering]
L --> F[Fleet Engineering]
| Layer | Question | Unit | Canonical reference |
|---|---|---|---|
| Prompt | What do we say? | One turn | — |
| Context | What does the model see? | One inference | Cobus — Context Engineering |
| Harness | How does one run execute? | One agent run | Cobus — Harness Engineering · awesome-harness-engineering |
| Loop | What keeps working over time? | One autonomous system | loop-engineering |
| Fleet | How do many systems govern together? | Population | This repo |
| Signal | Graduate to |
|---|---|
| Context window fills or rots | Context engineering |
| Single runs are flaky | Harness engineering |
| You are still typing prompts daily | Loop engineering |
| 3+ loops or 5+ agents; compliance questions | Fleet engineering |
- Loop repo → Fleet: loop-engineering multi-loop is the seed of fleet coordination
- Fleet repo → Loop: every registered agent should reference which loop pattern it runs
| Product | Layer |
|---|---|
| Claude Code, Codex, Grok CLI | Harness + Loop |
| LangGraph | Harness + Loop (orchestration) |
| LangSmith | Harness observability |
| LangSmith Fleet | Fleet (enterprise workspace) |
Fleet engineering as a discipline is bigger than any one product. Use the stack doc to place tools without conflating them.