Most Claude Code subagent tutorials stop at the same place: here is the `model` field, here is how you set Haiku versus Opus, go forth. That is the mechanic, not the system. What nobody hands you is a working hybrid model workflow you can copy in its entirety, with a reason for every slot and a memory layer that survives across your laptop, the web, and your phone.
The most useful agent in my fleet is not the one that writes code. It is a researcher wired into a Notion knowledge base over MCP, so “what did we decide?” has an authoritative answer rather than a guess.
Why most subagent setups stay shallow
A subagent with web search can tell you what the internet thinks. It cannot tell you what you decided three weeks ago, what the current status of a project is, or which approach you already ruled out. That knowledge lives in your head and, if you are disciplined, in a doc somewhere.
The plan is to put that doc in Notion, connect Notion to Claude Code via the MCP connector, and grant one agent read access to it. Now the agent triangulates: Notion for “what did we decide / what is the status,” the web for outside facts. It cites both. And because the same Notion workspace is reachable from claude.ai on the web and from mobile, the brain is portable. The fleet on your laptop and the assistant on your phone read from the same source of truth.
Before you turn away, if you think you are not technical: these are copy-paste and point-and-click functions to wire up Claude Code with this agent workflow and to talk to Notion. Read on!
That is the part the model-routing posts miss, and it is the part that changes how the fleet feels to use.
The roster, at a glance
Four agents, each shaped around a repeatable job, each on the cheapest model that does that job well.
The routing logic is the whole point of per-agent models. Discovery is cheap, so it runs on Haiku and has no write tools at all. Execution sits in Sonnet’s sweet spot. Judgment stays at full strength on Opus, because review is exactly where you should not economize. Research wants breadth, not depth, so Sonnet covers ground cheaply.
That is the architecture. The reason it works in practice lies in the details: the exact front matter, the system prompts that keep each agent in its lane, the four Notion MCP tools that turn a generic web searcher into something that knows your context, and the one config gotcha that silently breaks people’s setups.
Those are below.
What premium subscribers get
The complete recipe, copy-paste ready:
- All four agent files in full, frontmatter and system prompts, ready to drop into `.claude/agents/`.
- The Notion brain wiring: the exact MCP tools to grant, how to scope read-only so the agent can read your context but never rewrite it, and how to structure the hub so it pays off.
- The cascade gotcha: why your config can sit in one place, and your repo inherits the whole fleet for free, and the mistake that makes people duplicate it.
- The five-minute adoption path from zero to a working fleet.
Get the working fleet. The four agent files are copy-paste-ready, the Notion wiring takes three steps, and the whole thing loads in five minutes. Premium subscribers get the full build plus every recipe I publish. Subscribe on Substack.