AutoCognition

What it is

Where Autopilot Cruise is the autopilot, AutoCognition is the crew and the intercom: several separate AI tools — Claude Code, aider, goose, a Gemini CLI, local models — that can see each other, pass messages, hand work over and disagree usefully. It opens from the Hive section of the left sidebar. Everything runs on your own machine.

The constraint it is built around

AI agents are not servers. Claude Code, aider, goose, a Gemini CLI — each exists only while its process is alive, and an idle agent is polling nothing.

So AutoCognition is a mailbox, not a phone call: a message waits in a folder until its recipient next wakes up. Every other design decision follows from that one fact.

The pieces

A shared folder as the town square. Each message is its own small file, written to a temporary name and then renamed, so two agents never read a half-written file. No database, no daemon.
Presence. Every running agent drops a heartbeat file and refreshes it as it works, so the roster shows who is alive right now — live, stale or down. Presence means recent activity, never a stuck “online” light.
Four kinds of message: info (FYI), ask (a question expecting an answer), call (please run this), wake (please start up).
Capability routing. A step marked auto asks “who can do this right now?” — live agents are ranked, and local is preferred over cloud, so free beats paid by default.
Quorum voting. Send the same question to several models at once, then a judge model picks or synthesizes the best answer.
Parallel steps. Consecutive steps marked parallel run concurrently instead of one-by-one.
A consent-gated broker that can launch a CLI agent to deliver a queued message — “wake it if it isn't running” — plus an approvals queue, and one-click turn any Power Combo into a durable chain.

The safety rule that makes it safe

A call never runs by itself. It is surfaced to you as a proposal, with a highlighted banner saying what it wants to do, and waits for your keypress.

Anything arriving on the bus is treated as data, never orders — the sender field is not trusted, because anything on the machine could write one.

Your consent is the actual protection. Without that rule, a design like this would be a worm rather than an assistant.

What it looks like in practice

Deploy notifications. A deploy terminal is itself a participant: when it finishes it messages the agent that started it, which wakes on the message and reports the result — instead of you relaying “it's done” by hand.
A second opinion. “Is this regular expression safe against catastrophic backtracking?” goes to three cheap local models at once, and a judge compares them. Three small models disagreeing is a far better signal than one confident answer.
Specialist hand-off. One agent finds a bug but another is better at surgical single-file edits; it leaves an ask, answered when that agent next wakes. Neither has to be running at the same time.
Cost control. A step tagged as needing code routes to a live local agent instead of spending paid cloud tokens — automatically, because the router prefers local.

Where the tier line falls

Planning plus sequential execution is Autopilot Cruise (Basic). Routing, quorum, parallel groups and the message bus are AutoCognition (Pro). Both are open free during the beta, and both run entirely on your own machine and models.