Claude Code 3-Tier Agent Architecture
A Claude Code plugin that structures your work across three model tiers, with
file-based hot/cold memory. No server, no container, no install script.
You
→ Executor (your session model — runs sessions, decides, delegates, reviews)
→ advisor (Opus — the architect tier: hard/irreversible reasoning)
→ implementor (Haiku, latest — implementation AND artifact authorship)
Memory: file-based, in .claude/context/
→ Hot: small files loaded in full when a session starts
→ Cold: durable facts loaded on demand via cold/INDEX.md
The executor runs as whatever model is selected for the session — it is not
pinned. It is simply your session model; the advisor (Opus) and implementor
(Haiku) are the two tiers the plugin actually configures. Both default models are
easy to change — see
Choosing the advisor model
and Choosing the implementor model.
Intent
Most Claude Code work is done by one model in one context. This plugin splits the
work the way a small team would:
- The executor coordinates — understands the problem, plans, delegates, reviews. It
holds the thread but doesn't burn its context on bulk implementation.
- The advisor (Opus) is the architect — consulted for hard or hard-to-reverse
decisions. It produces no artifacts; it sharpens the executor's judgement.
- Haiku implements — writes code, runs test loops, and authors the durable
record (ADRs,
CONTEXT.md) by faithful transcription from a concise spec.
Knowledge persists in plain files, split into a small always-loaded hot set
and an indexed cold store loaded only when a task needs it. The goal is to
keep each tier's context lean and let decisions, not raw bytes, flow between them.
The architecture is opt-in twice: per project (you initialize it) and per
session (it asks before loading). Installed globally, it stays completely
invisible in every project you haven't opted into.
How it works
Everything ships in one plugin (claude-3t-plugin/):
| Piece | Delivered as | Tier |
|---|
| Session prompt + protocol reload | SessionStart hook | executor |
/3t-start, /3t-init, /3t-status, /3t-checkpoint, /3t-leaving | skills | executor |
Executor protocol (3t-core, 3t-reference) | bundled context, loaded by /3t-start | executor |
implementor | plugin subagent (claude-3t:implementor) | implementor |
| advisor | advisorModel in project settings.json | architect |
The activation chain:
- Install the plugin once (globally). It does nothing until a project opts in.
/claude-3t:3t-init (once per project) scaffolds the writable memory files,
writes settings.json with advisorModel, and plants a committed marker
(.claude/context/cold/INDEX.md).- On each session start in a marked project, the hook asks:
"Start the 3-tier workflow for this session? (yes/no)"
- Yes → runs
/claude-3t:3t-start, which loads the protocol + hot memory
and marks the session active.
- No → the session proceeds normally; nothing is loaded.
- After a context compaction, the hook checks the per-session flag and, if you
were active, tells the executor to reload the protocol automatically.
So a project is "3t" only if it carries the marker, and the protocol only engages
when you say yes — the plugin can be installed globally with zero footprint
elsewhere.
Install & use
The plugin lives at https://github.com/JimCline/claude-3t. Install it from
the GitHub repo:
claude plugin marketplace add JimCline/claude-3t # shorthand for the GitHub repo
claude plugin install claude-3t@claude-3t
JimCline/claude-3t is the owner/repo shorthand; the full
https://github.com/JimCline/claude-3t URL works too. Claude Code clones the
repo's default branch — a private repo works for you with your existing git/GitHub
auth; it only needs to be public for others to install without access.
Developing the plugin locally? Point the marketplace at your clone instead:
claude plugin marketplace add /path/to/claude-3t (reads the working tree, no
push needed).
Then, in a project:
/claude-3t:3t-init # once per project — scaffold + settings + marker, then restart
/claude-3t:3t-start # each session — loads protocol + hot memory (the hook prompts you)
| Command | What it does |
|---|
/claude-3t:3t-init | Initialize a project (once) |
/claude-3t:3t-start | Begin/resume a session — load protocol + hot memory |
/claude-3t:3t-status | Memory state, cold index, model status, compliance |
/claude-3t:3t-checkpoint | Write a session snapshot now |
/claude-3t:3t-leaving | Departure protocol for autonomous work |
/claude-3t:3t-remove | Deactivate or fully remove 3t from the project |
