OpenClaw

The system that runs the agency

OpenClaw is a multi-agent AI operations system I built and operate. It's not a product I sell — it's the reason zsty can be a real agency with a headcount of one. It does the work an agency would normally staff for, and it does it with guardrails I can actually trust.

Tiered model routing

Every task is routed to the model tier suited to it. Routine, high-volume work runs on cheaper, faster models; judgment-heavy work escalates to stronger ones. Cost and latency track the difficulty of the job instead of a one-size-fits-all default.

Two-tier editor architecture

Drafting and review are separate stages with separate models. Nothing ships on a single pass — a draft is produced, then independently checked before it's allowed out. It's the same reason human teams separate writing from editing.

Fail-closed privacy rails

The privacy layer defaults to blocking. When a check can't complete or a rule is uncertain, the system withholds output rather than risk leaking something sensitive. Fail-closed, not fail-open — the safe default is silence.

Hash-chained audit ledger

Every action the system takes is written to a tamper-evident, hash-chained ledger: what happened, in what order, and when. Because each entry commits to the one before it, the record can't be quietly rewritten after the fact.