de fabrica agentis

Anatomy of an AI Agent

Vesalius cut open the body and published his findings in 1543 so you could see how the parts connect. Now I am cutting open a production AI agent, so you can see how it actually works.

A Vesalius-style anatomical figure of an AI agent: a standing flayed body with eight labelled systems - the brain in the skull, the face, the nervous system, the muscles, the hands, the senses (eye and ear), the skin, and the mouth. Decision Boundaries VII Memory IV Orchestration Loop III The Model (LLM) I Output & Escalation VIII Tools / Connectors VI Skills V Identity / Persona II
Select a system to dissect it. Everything is readable without doing so.

The eight systems

  1. Cerebrum The Model (LLM) the model

    The reasoning core: the LLM that thinks. One organ, not the whole agent.

    model.md
    provider: claude-opus
role: reason over the task, propose the next step
  2. Facies Identity / Persona system prompt

    Who it is told to be: mandate, scope and tone, fixed in the system prompt.

    system.md
    You are a research assistant. You are precise,
you cite sources, and you never invent a fact.
  3. Nervi Orchestration Loop plan-act-observe

    The routine it runs every turn: load, plan, act, observe, then repeat.

    loop.md
    0 load context   1 plan
2 act (call a tool)   3 observe
4 draft output   → repeat
  4. Musculi Memory context store

    Muscle memory: what it kept from past runs. Plain files re-read each session, not a vector store.

    memory/
    notes.md      last read: this run
facts.md      re-loaded every session
# no vector store; plain files on disk
  5. Manus Skills sub-routines

    In the hands: routines it already knows, each with its own steps and its own limits.

    skills/
    - summarize_thread
- draft_reply
- extract_action_items
  6. Sensus Tools / Connectors function-calling

    Eyes, ears and a reach for outside systems: a search, a database, a file it writes.

    tools.json
    { "name": "search_web", "type": "function" }
{ "name": "run_sql",   "type": "function" }
  7. Cutis Decision Boundaries guardrails

    The split between what it decides alone and what waits for you. Autonomy made safe.

    policy.yaml
    can_send_email: false
can_spend: false
requires_human_approval: [publish, delete, pay]
  8. Vox Output & Escalation human-in-the-loop

    One channel back to a human. The agent drafts, you approve, nothing ships alone.

    output.md
    status: draft
destination: /drafts/
# a human approves before anything ships

The other seven systems

Ask where the intelligence lives in an AI agent and everyone shows the model. Fair: the model is the brain, the part that reasons. It is also one organ. A body is not a brain in a jar, and an agent is not a model with a prompt taped to it.

Look at the figure again. Seven of the eight systems are not the model. They decide whether a clever answer becomes work you can use: what the agent remembers, what it can reach, what it is allowed to touch, who signs the result. An agent that fails almost never fails at the brain. It fails at one of the other seven, and you can usually point to which.

Memory is the one that fools people. It is the muscles, and the tell is muscle memory: the agent is fluent at what it did last run and blank on what it did not carry in. No deep well feeds it. It re-reads a handful of files at the start of every run, acts sure of them, and forgets the rest. When an agent contradicts itself between Tuesday and Thursday, this is why. Not a weak model. A thin muscle.

Skills and tools look the same. They are not. A skill sits in the hands: a routine the agent already knows, with its own steps and its own limits. A tool is something it picks up: a search, a database, a file it writes. You would not confuse your handwriting with the pen. The agents that hold up keep that line sharp; the ones that wander are reaching for a tool when they needed a skill, or the reverse.

Guardrails are the skin. Most of the design work on a real agent goes into what it may not do: what it refuses, what it escalates, what it hands back before acting. Autonomy is a boundary drawn tight enough that the agent can run inside it, not the model let loose.

The mouth is the last organ, and the smallest. One channel out, a human at the valve. The agent drafts, a person approves, nothing ships on its own. That was designed in, not bolted on late.

You build an agent the way you build a body, one system at a time: add the organ that was missing, feed the one that was thin. The brain was never the hard part. The hard part is everything around it, the work you finish before you let it speak.

Dissected by Umut Ozbay, VlotAI.