Agent Anatomy: Brain, Tools, Memory, Policy

The model is only one organ

A strong model raises the ceiling, but the rest of the body decides whether the agent survives real work. Tools determine what it can do. Memory determines what it can carry forward. Policy determines what it must never do. Observability determines whether you can debug it after it disappoints you, because yes, it will.

The five-part stack

Brain: model selection, prompt, reasoning budget, output format.
Hands: tools, hosted tools, MCP servers, APIs, local shell, browser, code execution.
Memory: short-term context, scratchpad, vector recall, structured facts, durable notes.
Nerves: traces, logs, metrics, replayable runs, eval records.
Immune system: permissions, guardrails, sandboxing, human approval, kill switch.

You can replace one part without replacing the whole. That is why a good agent design is model-agnostic but not model-indifferent.

Design by blast radius

The more a tool can change, the tighter its policy should be. Reading a public page is low blast radius. Charging a card, deleting a database row, or pushing to production is high blast radius. The agent should feel that difference in code, not just in a hopeful prompt sentence.

Code

Capability map·yaml

agent:
  brain:
    model: "provider-selected"
    output_contract: "final_answer | tool_calls | handoff"
  hands:
    read_tools: ["web_search", "read_file"]
    write_tools: ["write_file"]
    destructive_tools: ["delete_file", "deploy"]
  memory:
    short_term: "conversation"
    long_term: "sqlite + markdown notes"
  policy:
    require_approval: ["delete_file", "deploy"]
    max_steps: 20
    max_cost_usd: 1.00

Agent Anatomy: Brain, Tools, Memory, Policy

The model is only one organ

The five-part stack

Design by blast radius

Code

External links

Exercise

Progress

Comments 0