The hybrid architecture pattern

A librarian doesn’t memorize every book in the library. She knows the catalog: what exists, and where. Ask her about maritime law, and she walks straight to the right shelf.

The catalog lives in her head. The books stay on the shelves. Nobody suggests she memorize the library; nobody suggests she work without the catalog. Both halves of that arrangement are load-bearing, and both have exact equivalents in agent design.

This chapter names the pattern that resolves Part 2’s central tension (every-turn cost versus available capability) and then shows you the same pattern hiding in three more places, so you’ll recognize it everywhere for the rest of the playbook.

The pattern, precisely

Whatever is most important goes in the system prompt, but for bulky content, what goes in the prompt is the metadata: the fact that the content exists, what it contains, and when to consult it. The real data goes in the referential store.

Anatomy of the prompt-side half:

## AVAILABLE REFERENCES (consult via knowledge search when relevant)
- "CBT Manual": cognitive-behavioral techniques; consult when the trainee
  asks about thought records or behavioral activation
- "Pricing Sheet 2026": current plans and discounts; consult for any pricing question
- "Objection Playbook": 12 scripted objection patterns; consult before
  running the objection drill

Three lines of catalog replace thirty thousand tokens of content. And the catalog fixes the failure mode of a pure knowledge-base setup, which is subtle and common: the agent can’t search for what it doesn’t know exists. Without the catalog line, the objection playbook sits in the store, perfectly retrievable, never retrieved, because nothing ever told the agent there was a playbook to consult.

Pure prompt fails on cost and dilution (Chapter 5). Pure store fails on discoverability. The hybrid keeps every-turn cost low while keeping every capability discoverable. That is the whole trick, and it is enough.

The same pattern, three more places

You have already seen this pattern once, and you are about to see it twice more. Cementing it now saves three future explanations.

1. Skill files in coding agents. Claude-Code-class agents keep one line of metadata per skill in the system prompt; the skill’s full body is referential, loaded just in time when invoked. Hundreds of skills, near-zero standing cost.

2. Agent memory (Part 3, next). A compact “super memory” is injected into the prompt (name, last session, preferences) while the full history is referential, searchable via a memory tool. Summary resident, detail on demand. That is the hybrid pattern wearing a different name, and Chapter 12 builds on it directly.

3. Sub-agent handoffs (Part 4). When one agent hands a session to the next, the handoff note it passes is metadata about the full session; the transcript stays behind, queryable if genuinely needed. Same split, third costume.

One pattern, four appearances. When you meet a new context problem this playbook doesn’t cover, the first question to ask is: what’s the catalog, and what’s the shelf?

Design procedure

A checklist you can apply to any resource, today:

1. List every piece of information your agent might need.
2. Sort by cadence (Chapter 6): every-turn goes in the prompt; per-session goes in the memory injection; occasional goes referential.
3. Write the catalog entry for each referential resource: one line covering name, contents, trigger condition.
4. Put the catalog in the prompt under a clearly marked section, like the block above.
5. Verify in transcripts: does the agent actually consult the resource at the trigger moments? If not, sharpen the trigger line. The catalog entry is a tool instruction, and Chapter 8’s rules apply to it, including the frequency line.

Example: the therapist trainer, before and after. Before: a 60,000-token prompt with five embedded manuals. Slow, expensive, and the coaching personality buried under protocol text. After: a 3,000-token prompt holding the role, the flow, and five catalog lines; the manuals in the knowledge base; retrieval landing at natural pauses per Chapter 9. Same knowledge available, an order of magnitude cheaper per turn. And, per Chapter 5’s accuracy curve, measurably sharper behavior. The hybrid pattern is one of the few changes that improves cost and quality in the same move.

The hybrid pattern’s most personal application is memory: what your agent remembers about the human it’s talking to, across every conversation they ever have. That is Part 3.