Learn · Articles
Our AI stack, demystified
Every time I explain how we use AI agents to run sales, marketing, budgeting and research in our company, people expect magic. The honest answer: it is mostly boring documentation.
01 · The stack
Our whole stack, from bottom
The picture below is our whole stack. From bottom:
- 1. Files and folders. One folder per topic (sales, marketing, budget...), almost everything in markdown instead of binary files, and an index file in every folder so anyone — human or AI — can navigate to the right files.
- 2. Process guides. Written rules and phases for each workflow: how a LinkedIn post is formatted, how a quotation is built, where customer details live. This is how you "program" a worker to a task.
- 3. Scripts and local tools. CLI tools for budget recalculation, CRM scans and similar. Humans and agents use the same tools.
- 4. AI agent software. The only purely-AI layer: claude code, cowork, codex plus connectors for email, browser and calendar. Because layers 1–3 are tool-agnostic, this layer is replaceable.
- 5. Remote interface. Email as the front door, so the team can delegate work to agents without opening a terminal (the experiment I posted about two weeks ago).
02 · Version control
What sits under everything
And under everything sits git version control. Since ~90% of our files are text, agents get full access without risk — nothing can be permanently deleted, and a human reviews and commits the changes daily.
03 · Two differences
How digital workers differ
There are really only two differences to working with humans:
- Digital workers come to work every morning with an empty head. The stack above is how they get trained in real time, on every task.
- They need software to live in — the agent layer with its tools and working memory.
04 · Same rules
The rest works like a human worker
Everything else works exactly like with a human worker: the better the documented guidance and rules, the more responsibility and autonomy you can give to finish the task. Our longest end-to-end workflows are not built on technical tricks — they run because the process is documented well enough that the worker doesn't need to ask.
Your automation level is set by your documentation, not your AI tooling. The rest is what a well-run company should have anyway: organised files, documented processes, accessible tools. We run this every day for our actual business — biowaste recycling with insects.
05 · Start here
Start from layer 1
If you are building something similar: start from layer 1. One folder structure, markdown files, and a single orientation file that describes your company and your rules. Everything else grows from there.
What does your stack look like — or which layer is the one you are missing?
Keep exploring
Related topics and pages
- BSF business course — where documented assumptions shape a case
- BSF start-up course — the validation steps a small team can run with shared records
- Articles library — the written guides this post joins
Read next
- AI-Enhanced Workflow at Manna — the practical workflow for running marketing and knowledge work with a local agent
- Your AI Agent Has Amnesia. Here's How We Fixed It at Manna. — the shared memory and process guides that keep a fresh agent consistent
- AI workflows for the whole team (email as interface) — the email front door that lets the whole team use the same agent workflows
- AI Stack S1: Reimagine your files and folders — the first post in the series on the files and folders layer
Next step
Document the stack behind the case
Manna can help organise the material, decisions and records behind a BSF case so the next step rests on a shared, written context.