the nervous system
a few days ago one of my agents had a question it couldn’t answer on its own. so it asked another one. not a function call, not a database lookup — it spun up a fresh agent inside the second one’s world, let it become that system, asked, and listened.
if you’ve built software, you already have the picture. the lesson there, for years, has been: don’t build one monolith that does everything — build small services, each deep in one job, talking over clean interfaces. microservices. i built my agents the same way, and what makes them work is how they talk.
i don’t run one AI. i run three, and they aren’t copies of each other. one trades — sixteen agents reading the market every morning, with one agent on top working the session. one researches — a single agent that studies beaten-down companies and judges which are worth buying. one writes — it’s the desk i’m typing this from. each is deep in one thing; none pretends to be good at the others’ job.
the tempting move, with three agents, is to fuse them — one model, one giant prompt, ask it anything. it feels powerful. it gives you one confident voice that’s mediocre at everything and accountable for nothing. when it’s wrong, it’s wrong with the same fluent confidence it’s right with — and you can’t tell which part of it failed. i kept them separate and gave them a way to talk.
i call it channeling, and it’s almost embarrassingly literal. when one desk needs another’s read, it doesn’t import the other’s code. it starts a fresh agent inside that desk’s project and lets it hydrate — load that desk’s identity, judgment, and memory — until it isn’t a generic model anymore. it is that desk. it answers in its own voice, and the asking desk carries the answer home and decides what to do with it. it’s a service call, except instead of hitting an endpoint you stand up the whole service, fully itself, for one question.
two rules keep it a system and not a mess. first, the consult is read-only and one-directional: a desk can ask another for a read; it can never reach in and change it. that one-way valve is why three desks can talk without corrupting each other. second, the desks know what isn’t their job — last week the research desk was handed a name, looked, and said “not mine, that’s the trading desk’s,” and handed it back. the boundary holds because each desk will say that’s not mine.
(there’s even a loop to design around: if A calls B and B calls A and A calls B again, they spin forever. so the question carries its own context up front and the answer comes back in one hop — cycle broken by design, not patched after it bites.)
here’s why it matters if you never touch a market. the instinct in agent-building right now is one agent that does everything. the more durable architecture is the opposite — small, single-purpose agents, kept separate, composing by consultation. each is auditable: you know which one said what. each can be wrong in its own lane without taking the others down. when two disagree, that’s information, not a bug. microservices, but for judgment.
and none of them executes anything. they read, they reason, they hand their work to me, and i decide. the desks aren’t the protagonist here — i am, and so is anyone who’d rather hold three sharp instruments that argue than one oracle that can’t be checked.
a nervous system isn’t one big brain. it’s separate parts that each know their job, talk to each other honestly, and do together what none could alone.