Chapter
Climb the eight tiers — to spend effort only where your work needs it
This guide is one ladder: eight tiers, simple → hard. A tier is a level of skill and a level of where you're applying effort.
As you climb, the work shifts from wording the prompt to engineering the system around the model — prompts shrink, the system carries the intelligence.
You don't need all eight. Climb until the agent is as reliable as the work demands, then stop.
The trick is knowing what pushes you to the next tier — it's always a specific pain, not ambition. When you feel the pain in the right column, you're ready for the next tier.
| Tier | What you work on | Level it applies at | What pushes you up |
|---|---|---|---|
| T1 — Professional Prompting | The single request | One message | The agent keeps doing almost the right thing — vague asks get literal, wrong results |
| T2 — Shaping & Slicing | The task before you start | One task | Big asks go sideways; it edits the wrong things or tries to do everything in one pass |
| T3 — Context Management | The project the agent sees | The repo | You re-explain the same conventions every session; it can't see your DB/browser/docs |
| T4 — Loop Until Done | A "done" the agent can check itself | The task, automated | You can't trust the output without reading every line; "done" means nothing concrete |
| T5 — Checkpointing & Hardening | Safe, revertible runs | The session | A long run goes wrong and you lose good work; nothing to roll back to |
| T6 — Orchestration | Many subagents, long horizons | Multi-step / multi-agent | One agent is too slow or floods its context; the build is too big for one pass |
| T7 — Fleet Ops | Where & how runs execute | Your machines | Runs die when your laptop sleeps; parallel agents collide; you want to drive from your phone |
| T8 — Agent Execution Layer | Agents as async workers | Your org / production | The team needs it: agents must pick up tickets and open PRs without anyone babysitting a terminal |
Read the right column as a diagnostic. Stuck re-typing the same context every session? That's the Tier 3 pain — go engineer CLAUDE.md and Skills.
Losing work on long runs? Tier 5 — commit on every green. Each tier exists to kill a specific failure of the one below it:
Prompts drift → T2 plans them. Plans are forgotten across sessions → T3 makes context durable. Context still can't prove correctness → T4 verifies. Long runs lose good work → T5 checkpoints. One agent is too slow → T6 orchestrates many. Runs die when your laptop sleeps → T7 runs a fleet. Humans still babysit terminals → T8 makes agents async workers.
The destination, if you go all the way, is professional agentic product engineering: agents running in a loop against a clear, testable standard, inside a real repo you own.
Which tier do I need?
Match the job to a target tier and stop there — climbing higher than the work demands is wasted effort.
| What you're building | Aim for | Why |
|---|---|---|
| A landing page, throwaway script, or one-off for yourself | Tier 1–2 — or skip the agent and use a one-shot builder (Lovable/v0/Bolt) | Low stakes, short-lived. A clear prompt is enough; you don't need tests, git discipline, or a loop |
| A feature in a side project you intend to keep | Tier 1–4 | Now correctness matters over time — you want a verify loop and checkpoints so it doesn't rot |
| Production code in a shared team repo | Tier 1–5, reaching into 6–8 as scale demands | Multiple people, real consequences. Context engineering, git hygiene, and review become non-negotiable; orchestration and a hosted execution layer follow when one agent or one machine isn't enough |
So for that simple landing page: Tier 1–2 is the answer — and a one-shot builder may serve you better than Claude Code (see Agents vs one-shot builders).