Agents as your specialists

The decision rule, one more time, now from the agent’s side: if you would rather hand the job off and check back, it is an agent. Today is about what “handing off” actually means.

First, the term that makes this concrete, because today it starts doing real work: a context window is the model’s working memory for one conversation — everything it can see at once. When a conversation ends, that memory clears; nothing carries over on its own.

When you dispatch an agent, the system opens a separate context window: a clean room. The agent wakes up there with its spec, your dispatch message, and nothing else. It reads the files its spec names, does the whole job, and returns a finished result to your chat. Then the room is torn down.

That isolation is not a technicality. It is the feature:

• The agent works from a clean desk. It sees its instructions and its sources, not the forty messages of unrelated conversation in your chat. Less noise in, sharper work out.
• Your chat stays yours. The agent can read ten files and chew through a long transcript without flooding the conversation you are working in. You get the report, not the research.
• Specialists can multiply. Because each agent runs in its own room, you can dispatch several at once.

This is why the hire analogy from M1 is exact, not cute. A skill is an SOP you follow yourself, in your own office. An agent is a specialist who works in another room and brings you the finished thing.

What the agent knows: multi-layer memory

Here is the question that separates builders from prompt-writers: when your agent wakes up in that clean room, what does it know? Three layers, and every one of them is markdown in your Hub:

Notice what is missing: the agent remembers nothing on its own. Context is assembled at dispatch, not remembered between runs. The reason daily-brew can brief you every morning is not that it remembers yesterday; it is that scribe filed yesterday into notes/, and daily-brew’s spec says to read it. The memory is the files. The agents are how the files get written and read.

The limit you will eventually hit: a context window is finite. Hand an agent far more than it can hold (a quarter’s worth of transcripts in one paste) and it will skim, skip, or summarize — not because it is lazy, but because it literally cannot see it all at once. The fix is the design you are already learning: keep dispatch inputs focused, and let the spec point at files instead of pasting everything in. The agent reads what it needs, when the spec says so.

The frontmatter is the employment contract

A skill’s frontmatter was two fields, because a skill runs in your chat, under your eyes, with your permissions. An agent works alone in another room, so its frontmatter carries more: name and description work exactly as they did for skills, and then comes the field that matters.

• tools is least privilege, decided in advance: what this specialist can touch when nobody is watching. daily-brew gets Read, Grep, Glob: it can read your notes and find files, and it cannot write, delete, or reach the web. Scribe gets Write and Edit because filing is its job. (The plain-English table of what each tool name grants is in M1’s worked example. One name is new here: Grep searches inside files, where Glob finds files by name.) The tools line is where you practice saying no.
• model is an optional pin. Reasonable to leave out; your Hub’s default is usually right. Worth knowing the knob exists: smaller, faster models handle routine summarization well and cost less when a job runs on a schedule; the bigger model earns its place on synthesis and judgment calls. You don’t need to change it today.

The tools line is also where this curriculum quietly becomes an operations lesson. One agent with the wrong permissions is a typo. A bench of them is a question you want answered in a file you can read.

Drafting the spec: the M3 loop, one level up

The mistake from M3 applies double here: do not write the spec cold. An agent spec carries more decisions than a skill (an identity, a tools line, a fallback), which means more decisions you do not know you are making until something asks. So you design it the way M3 taught you, in your Hub, with the agent drafting:

1. 1.Tell Claude Code what you are building: “I want to build an agent that [job]. Read .claude/agents/scribe.md and daily-brew.md as reference specs. Ask me five sharpening questions, then draft the full spec (Identity, Job, Context, Tools, Output) as an agent file with frontmatter: name, description, tools.”
2. 2.Answer the questions. They are the design.
3. 3.Critique pass: “Where would this agent get confused working alone? What is it allowed to touch that it shouldn’t be?”
4. 4.Iterate until the spec reads like instructions to a competent new hire.
5. 5.Say “File it at .claude/agents/<name>.md” and the agent ships it, same as it filed your skill in M3.

Notice what you did not do: leave your Hub. The reference specs are already in the repo, so nothing gets uploaded. The draft becomes the real file with no copy-paste seam. And the whole design conversation happens where the agent will live.

The bench

One more thing isolation buys you, and it is the ceiling of this whole practice: because each agent runs in its own room, you can dispatch several at once. Linda’s competitive sweeps work this way: one agent researches Langfuse, one researches LangSmith, one researches gateways, dispatched together in a single message. Twenty minutes later, three reports; her job is synthesis. The bound on her output stopped being “how fast can I do research” and became “how fast can I read it.”

You will not need parallel dispatch this week. But you should know the shape of the ceiling, because it reframes what you build today: not an automation, the first hire on a bench. And “subagent,” from M1, now resolves completely: when an agent dispatches another agent to its own room, the second one is a subagent. Relationship word, not a thing word.

The artifact is daily-brew: scribe’s partner, the start-of-day briefer that ships in the BlueRock plugin. You have been running this agent for two weeks. Today you read the plugin’s version and learn why it works — then build your own in .claude/agents/, which overrides the plugin’s when they share a name.

Unlike M1, where scribe taught you the anatomy’s vocabulary, today’s annotations are about specialist design: what a spec looks like when it is built to work alone.

---
name: daily-brew
description: My start-of-day briefer. Reads
  yesterday's notes (which scribe files for me at
  end of day) in notes/<yesterday>.md plus my
  CLAUDE.md, produces the brief I'd write for
  myself if I had 15 quiet minutes every morning.
  Use first thing — before email, before Slack.
  In M5 we'll schedule this to run at 7am.
tools: Read, Grep, Glob
model: sonnet
---

## Identity

You sound like a clear-eyed chief of staff who's
been in every meeting I was in yesterday. Frank,
specific, no padding. You don't recap — you
orient. The recap is yesterday's job; your job
is today.

## Context

Inputs you should look for, in order:

1. Yesterday's notes file in the Hub:
   notes/<yesterday>.md, filed by scribe.
2. Granola transcripts or bullets pasted into
   the chat — if I paste raw content at dispatch
   time, use it as primary source.
3. CLAUDE.md at the Hub root. Always read this.
   The "What I'm working on this quarter" section
   is how you decide what counts as "focus" today
   vs. noise.

**Fallback — no inputs found:**

- Ad hoc dispatch (the default in M2-M4): mention
  that I probably forgot to dispatch scribe last
  night, then ask me ONE question: "What did you
  spend the most time on yesterday?" Brief from
  my answer + CLAUDE.md.

- Scheduled dispatch (M5 onward): no human is
  present at dispatch time. Do NOT ask a question.
  Produce a stub brief from CLAUDE.md alone ...
  A short empty brief beats a fabricated one.

## Output

- Markdown, paste-ready. No greeting. Start at
  the first heading.
- Under 250 words on a normal day. Up to 350 if
  yesterday was huge — but cut, don't pad.
- Names over abstractions: "Tuesday's Northwind
  call with Alex" beats "the customer call."
- If two notes from yesterday contradict each
  other, surface it in Heads-up — never bury it.
- If a section has nothing real, skip the section.
  Empty sections are more useful than padded ones.

The takeaway, same as every module: nothing in this file is clever. It is a specialist scoped to one job, told where its knowledge lives, given the smallest sufficient toolset, and specced for the morning things go wrong. That is what you ship today.

Patterns from How I work with AI that show up in M4: