I run six AI agents that work while I sleep. Each one has a name and one job. One plans my day. One builds the others. One finds people I should know. One ships internal tools. One runs client accounts. The newest one gives second opinions.
They send each other signed messages, so a fake note cannot fool them. They all read the same shared memory. Together they are my back office.
The roster
Cairo is the chief of staff: morning brief, evening review, weekly retrospective, with a durable scheduler and typed subagents under a single-writer rule. Forge is the one that builds the others. Juniper researches and nurtures an advisor network. Kestrel ships and maintains internal tooling. The fifth agent's codename stays private. It is the account manager, running each client account on its own clock, with durable per-account scheduling, reconciled retries when a turn fails, a derived-state ledger per account, and an explicit policy for which emails are urgent enough to escalate. A sixth, Claude-based agent joined as the fleet's second opinion. Each runs as a single-replica service on Railway with its own persistent volume; about 16 GB of live agent state sits across those volumes.
How they talk
The agents talk full-mesh over private networking using Ed25519-signed envelopes with receipts, expiry, and recipient verification, and every injected peer message is explicitly labeled as not-from-me and carrying no human approval. Every agent holds credentials to Thoth, a shared long-term memory behind the whole fleet.
How new agents are born
To create a new agent, Forge interviews me, adapts a clean clone of itself into the new role, proves a local test gate, then provisions a fresh private repo, service, and volume for the peer and records it in a fleet registry. Peers never share credentials or infrastructure. Forge also carries a detached adversarial-review harness that runs a second model in a scrubbed environment to attack the fleet's own plans.
The rules they run under
The newest runtime draws the strictest boundary in the fleet: it refuses every path that would silently convert subscription usage into API billing. Setup tokens, API keys, gateways, and custom base URLs are all rejected, and the process exits rather than run in a state it should not be in. Each agent gets exactly one volume and one login, never copied from a sibling. Identity files seed once at first boot and are never overwritten after creation, so redeploys replace the code but never the agent's memory of who it is.
The liveness receipt
Over the past week, every service that ran the full week returned 10,080 of 10,080 one-minute metric samples, with no service dropping out.
Questions about this system, or the problem it could solve for you?
Discuss your project