// FIELD MANUAL · v2026.1
I build AI systems that remember who they are.
Multi-agent setups with persistent memory, running Claude Code at agency scale — 17 named agents, ~136 skills, six memory backends wired together.
- Persistent multi-agent systems. Agents that wake up knowing last week's work, not starting from zero every session.
- Claude Code at agency scale. 17 named agents with their own souls, workspaces, and routing — not one fat assistant pretending to be a team.
- Memory that actually holds. Hindsight, pgvector, graph stores, and doc stores working together, with one source of truth instead of six voices shouting over each other.
- Done-for-you automation. Real systems, not demos — the kind of plumbing that runs in the background while you're doing something else.
// THE CHEATSHEET
Claude Code 2026 — What a mature setup actually looks like
Most Claude Code setups top out at a bloated CLAUDE.md and a folder of half-used slash commands. That's month-one. This is what year-twelve looks like — the layers you build once you stop treating the assistant like a chatbot and start treating it like infrastructure.
// THE 9-LAYER ARCHITECTURE
L1
// LAYER 01 — FOUNDATION
CLAUDE.md files
Static context, layered global → repo → scope. Identity at the top, project state in the repo, nothing duplicated. If you're pasting the same paragraph in three places, you've already lost.
L2
// LAYER 02 — FOUNDATION
Skills
Routed behavioral patterns. The description field IS the wiring — that's how Claude decides to load it. Body is behavioral rules only, not lore. Audit them. Kill the ones nothing routes to.
L3
// LAYER 03 — PLUMBING
Hooks
Session lifecycle and memory plumbing. This is the nervous system — session start injects context, subagent start loads souls, tool events log provenance. Not a safety net. A signal path.
L4
// LAYER 04 — EXECUTION
Subagents
Single-task workers. One-way — they report back to the caller and die. Disposable unless you set memory: user in the frontmatter, which wires them to persistent memory across sessions.
L5
// LAYER 05 — EXECUTION
Agent teams
Peer coordination with two-way messaging. Experimental, gated behind a flag. Different beast from subagents — teammates are full sessions, not workers. Most people conflate the two and get burned.
L6
// LAYER 06 — IDENTITY
Per-agent workspaces
9-file OpenClaw pattern: AGENTS, SOUL, IDENTITY, USER, TOOLS, MEMORY, HEARTBEAT, BOOT, BOOTSTRAP. Each named agent gets its own. Identity lives in SOUL. State lives in MEMORY. Never mix them.
L7
// LAYER 07 — STATE
Memory backends
Vector stores, graph stores, doc stores — all working together. Hindsight for working memory, pgvector as source of truth, graph store for relationships, doc store for full text. Pick your roles before you wire anything.
L8
// LAYER 08 — OBSERVABILITY
MCC / dashboards
Mission control. Visibility across every agent, every session, every memory write. If you can't see what your agents are doing, you don't have a team — you have a prayer.
L9
// LAYER 09 — OBSERVABILITY
Provenance / audit
Cross-session trail. Every tool event logged, every memory write traceable. The forensic layer — the thing you'll need the first time an agent does something weird and you have to figure out why.
// FOUR ORGANIZING PRINCIPLES
01
CLAUDE.md is a lean index, never an encyclopedia
Global is identity only. Project is project state only. The moment CLAUDE.md starts explaining how your codebase works, you're doing it wrong — that's what skills and reference files are for. Lean index, pointers out, nothing duplicated.
02
Skills are classified, audited, and trigger-word disciplined
Frontmatter IS the wiring. The description field is how the router finds the skill — vague descriptions mean vague routing. Body is behavioral rules only. Audit the pool on a schedule and delete the ones nothing triggers. A dead skill is worse than no skill.
03
Hooks are the nervous system, not a safety net
Most people discover hooks when they want to block something. That's the small use case. The real job is plumbing — session lifecycle, memory injection, provenance logging, cross-system bridges. Build the hooks first and the agents get smarter for free.
04
Memory uses multiple backends — exactly one is source of truth
Multiple stores is fine. Multiple stores injecting into the same recall path at the same time is chaos. Pick one as the source of truth. The others are caches, backups, or specialized indexes. Never let two memory systems argue in front of the model.