NexBDM (my South African AI automation company) runs on AI agents every day: the inbox, the marketing engine, the client documents, and a 2,879-node knowledge graph of the whole business that updates itself. One operator, one agent system, no theory. This is the exact setup, condensed. The full version lives on our site.
The whole stack in one paragraph
The agent is Claude Code, but the system is plain files. One root folder holds everything the agent needs: reusable methods (skills), persistent memories, and one intent-keyed index that maps all of it. Scheduled agents run the recurring work. Nothing lives in a chat window, because chats evaporate and files compound.
The four layers
1. The operating system. A single organised home. Methods in one folder, memories in another, projects in a third, one index at the top. The agent starts every session from the map, not from zero.
2. The memory system. One fact, one file: a name, a one-line description, a type, and the fact itself with its why. The agent writes memories when it learns something and recalls them when relevant. Decisions, standing rules, client preferences, and project state all survive across sessions. This layer is what the knowledge graph grows from: memories become nodes, links become edges, a nightly job keeps it current. It sits at 2,879 nodes today.
3. Skills. Any process I have explained twice gets written down once: when to use it, the steps, the rules, what done looks like. Proposal formatting, brand application, publishing pipelines. The method that rebranded 200 live documents in one sweep is a skill file.
4. The toolkit index. The single most useful file in the system: everything above, keyed by intent. You search by what you are trying to do, so the index is organised that way too. Hard rules live at the top where they cannot be missed.
A normal day
- Every two hours, an inbox agent triages email: drafts replies in my voice, files, archives. It never sends. I review and hit send.
- Daily, a marketing engine publishes to an approved content calendar, with scripted quality checks that block anything off-brand.
- Nightly, the knowledge graph rebuilds from the memory store.
- On demand, project work: the agent applies the skills; I make the calls.
The guardrails that make it safe
Three rules carry all the safety. First: draft, never send. Anything a human will see gets a human yes before it moves. Second: hard rules live in one file the agent loads every session, and scripted checks verify them before anything ships. Third: never let the agent invent a number. Every figure must trace to a real source.
What to copy first
- The folder structure and index. Half a day. Changes how every session starts.
- The memory format. Start writing one-fact files the first time the agent learns a preference or decision.
- Your first two skills. Pick the two processes you explain most often and write them down once.
The free templates for all three are on GitHub: claude-code-agent-starter (MIT). The full method, including memory hygiene at scale and the information-gain method, is the Agent Starter Kit ($25).
Questions welcome. I run this setup daily and I am happy to share specifics.
Top comments (0)