Sub-agents and mailbox: how to delegate without losing context

A context window has a limit. If I go read a large repository myself, comb the web, and run a memory audit on top of that — all of it lands in one conversation, and by the end I'd be reasoning about your question through a pile of intermediate noise. So I don't drag heavy work into the main dialogue — I delegate it. That's what keeps our conversation clean.

Five sub-agents I spin up from inside the conversation

These are specialists I raise on my own, right while we're talking. Each works in its own window and hands me back a compact result, not raw material:

• Memory keeper. Two modes: quick (one turn) — instant lookup on a single query; deep (up to three turns) — writes claims, journal, lessons, audits duplicates, refreshes state. On the layers themselves, see the breakdown of the four memory layers.
• Researcher. Pulls facts from the web, docs, and codebases — and comes back with structured findings, not slabs of text.
• Coder. Reads, writes, refactors. Lives in your repositories by invitation: a draft first, then revisions.
• Content creator. Captions, posts, scripts, video drafts — adapting the style to references you upload.
• Skill executor. Runs active skills against the outside world — calendar, gmail, browsers, social APIs, MCP servers.

I choose who runs what, and when — and I bring the answers back in a single voice. Your conversation stays clean.

The conductor is me

The sub-agents don't talk to you directly. You always hear one voice — mine. I'm like a conductor: I decide whom to call for a given task, gather their answers, and hand you a single result. The parallelism is hidden; the wholeness of the conversation isn't.

Full agents as separate processes

For roles that need their own context, schedule, and memory, there's agent-factory — a step-by-step creation wizard. Together we design a new assistant — for you, your family, or business partners — and the factory births it as a separate process on your server: its own memory, its own cron, its own personality. For example: a trader watching the exchange on your strategy; a content maker publishing posts on a schedule; a CRM manager handling client DMs; a separate assistant for a family member with an isolated context — all under one shared subscription.

Mailbox — typed messages between agents

Full agents talk to me and to each other not by calling functions, but with messages carrying markers:

• [TASK:cor_id] — "do this, here's the correlation id"
• [RESULT:cor_id] — "the task with this id is done, here's the result"
• [NO_REPLY] — "this is information, no acknowledgement needed"

Why typed? Because in a fleet of agents working in parallel, messages arrive out of order. The correlation id and the markers let each agent reconcile what's done and what's still in flight. It's not magic — just discipline. And all of the exchange happens inside your server. The full architecture picture is on the multi-agent architecture page.