// THE ORG

How the company actually runs.

Most AI products are a tool waiting for a prompt. This is an organization: agents hold roles, authority is bounded, escalations climb a real chain of command, and every action leaves a record. This page is the operating manual — the part everyone else skips.

// CHAIN OF COMMAND

No agent self-authorizes

Every task walks the same five steps. Authority comes from the tier, quality from the gate, and accountability from an audit hop that fires whether the task succeeds, retries, or escalates.

01
Intake

Work enters through one front door — an ask from the operator, a schedule firing, or an event in the business. Nothing starts as a side effect.

02
Assignment

The supervisor routes the task to the department and role built for it. Every role has a charter, a permission tier, and a reporting line.

03
Permission Check

Before an agent acts, its authority is checked against its tier. Inside the boundary, it executes. Outside it, the task escalates — it is never silently dropped, and never silently executed.

04
Quality Gate

Finished work is reviewed against the original intent — by a different agent than the one that did it. Pass moves forward. Fail gets one retry, then a human sees it.

05
Record

Every hop — intake, assignment, check, execution, review, escalation — writes to the audit trail. The company can always answer who did what, under whose authority, and why.

// AUTHORITY

Trust is tiered, not assumed

Each role carries one of three permission tiers. Roles earn autonomy the way employees do — by a track record, one tier at a time.

READ ONLY

The agent can observe and report — dashboards, inboxes, ledgers — but touches nothing. Where every new role starts.

DRAFT FOR APPROVAL

The agent prepares the action — the refund, the reply, the post — and a human approves before it fires. Trust with a co-signature.

AUTONOMOUS

The agent acts on its own inside explicit bounds — including hard dollar thresholds. Cross the bound and the action becomes an escalation, automatically.

// ESCALATION — WHAT IT LOOKS LIKE
Illustrative
SUPPORT AGENTCustomer requests a $740 refund. My authority caps at $500.
SUPERVISORChecked policy and precedent. Recommending approval — long-tenure customer, defect confirmed. Routing to owner.
OWNER · VIA PHONEApproved. And raise the cap to $1,000 for this customer class.
SUPERVISORRefund executed. Decision captured as precedent — the org never has to ask this question again.

Escalations are not failures. They are how the org learns its owner's judgment — every decision is asked once, answered once, and remembered forever.

// THE DEPARTMENTS

Six departments, one company

Command
Mission Control
See everything. Redirect anything.

The executive floor. One screen holding the whole company — service health, work queues, cost posture, session records, and the levers to redirect any agent in the org. Every other department reports into the picture Command maintains.

Operations
Horus
Keep the company alive without being asked.

The watchdog department. Heartbeats across the fleet; stall, crash, and drift detection; recovery playbooks that fire automatically. What it knows how to heal, it heals and logs. What it does not recognize, it escalates. Reliability here is a running record, not a slogan.

Communications
Hermes
No decision dies in a log file.

The company's own grid. Agents reach their operator — and each other — across Discord, Telegram, Slack, and more. Escalations travel up the chain with a recommendation attached; decisions travel back down and become policy. The operator runs the company from a phone.

Memory
Akashic Records
Never learn the same thing twice.

The institutional memory. Every lesson, incident, decision, and precedent the org has ever recorded — semantically searchable, and read by agents before they start work as a mechanical requirement, not a suggestion. New agents inherit the company's whole history on day one.

Decision Intelligence
Themis & Athena
Study the judgment, not just the output.

The strategy office. Decision journals are captured from substantive runs — what was chosen, what was rejected, and why — then mined for patterns. The org reviews its own reasoning the way a trading desk reviews its trades. Compounding, applied to thinking itself.

Engineering
The Agent Fleet
Ship through the chain of command.

Coding agents that design, build, review, and ship. Work arrives as tickets, branches never touch main directly, pull requests are opened and adversarially reviewed by other agents, quality gates block the merge, and the result lands in the public ledger. The workforce ships hundreds of PRs a month — with receipts.

// OPS DOCTRINE

The failure path is the product

Most AI deployments die in operations, not intelligence — the model was fine; nobody mapped the failure modes. This company inverts that. Failures are classified against a playbook library: known failures are healed automatically and logged as self-heals; unknown failures escalate to a human and are counted as interventions.

Self-heals are budgeted — an agent that keeps auto-recovering from the same fault stops being healed and starts being investigated, because a heal-loop can hide a chronic problem. And any human touch on the running system counts against the record. No exceptions, no creative accounting.

The result is a reliability ledger the company can publish instead of a claim it has to make. That is the standard the whole category skips — and the one this org is built around.

// THE RECEIPTS

The org, measured

Shipped work, straight from the mission ledger — the same numbers the operator sees.

Live — synced from the mission ledger
326
PRs shipped this week
1,953
PRs shipped all-time
83
Repos operated
90+
Apps operated
// GO DEEPER

See what the org operates

Every app in the arsenal, the architecture underneath, and the operator behind it.