Can an agent organisation improve itself while you sleep?

Hypothesis

One agent with many tools plateaus. An organisation of agents — a chief of staff who only delegates, specialist officers, a reserve library of personas, work dispatched through a kanban board — should handle broader work, stay auditable, and improve itself on a schedule.

What we built

An autonomous org on our own server, built on the open-source Hermes agent framework: CEO (human, via Telegram) → chief of staff → five officers (science, engineering, comms, navigation, operations) → a reserve of 72 summonable specialist personas. Work flows through a kanban board, not chat. A nightly cron reviews the org's own performance and proposes improvements. Everything high-stakes passes a code-level approval gate with a human one-tap grant.

Learnings

• Delegation must be the default, not the fallback. We had to remove the chief of staff's own research tools before it stopped doing the work itself — org design for agents is mostly subtraction.
• An adversarial review of our own approval gate found five real bypasses that a static read-through missed. Never ship a security control without an attacker actually executing bypass attempts against it.
• A restore drill caught that our backups had silently excluded every database for weeks. Backups you haven't restored from are hypotheses.
• Persona files load on every call, so their size is an attention tax — identities stay lean, shared lore lives in the workspace.

Log

• 2026-07-11 — Hardening wave: two-tier approval gate, adversarial review (5 bypasses fixed), offsite backups, restore runbook, injection canary drill passed.
• 2026-07-08 — Org launched: chief of staff + officers + reserve library, kanban dispatch, delegation-by-default routing.