your digital twin
Philip is a typed-verb agent that lives inside the firm's existing primitives — identity, container platform, image trust, API gateways, inference clouds, immutable audit. Same shape Apple ships for itself.
tell me what's in my inbox · draft a reply to my latest customer mail · summarize the seven safe-agent layers · queue a daily morning brief at 7:30am · …
about Philip
Most enterprise AI today is a friendly receptionist with no permission slip — they accept any request, route it anywhere, and trust the model to behave. That works until the model is asked to do something it shouldn't.
Philip is the opposite. It's a typed-verb agent runtime where every action has to clear seven independent gates before it runs: identity, daemon entitlements, tool dispatch, user consent, model selection, tenancy envelope, and audit. Skip any one and the request is refused.
Same architecture Apple ships in intelligenceflowd, built end-to-end on the firm's own primitives — no new vendor, no new identity provider, no new SIEM. The Reference Architecture is documented end-to-end as a building you walk through, lobby to penthouse to vault.
the building
Imagine a sixty-story tower on Park Avenue. The lobby has security guards, a fingerprint reader at the turnstile, a concierge desk, a directory board, and an approval desk before the elevators. Working floors mid-building. A penthouse executive lab on the sixtieth floor where the hardest problems get worked on, sealed away. And below the lobby, behind a heavy steel door, the vault — where every entry is recorded forever.
the claw
A skill is a typed verb in Philip's tool registry. Concretely it's a Python function with a JSON-schema argument signature. The agent can chain skills to fulfill an intent — read your mail, summarize it, draft a reply — but every skill in the chain is independently checked. Manifest. Per-feature allowlist. User consent for sensitive actions. Six-axis tenancy envelope. Audit ledger.
Adding a new skill is a pull request: write the Python function, add it to the relevant feature's allowlist, update the manifest if it touches a new resource category. Engineering reviews; Trust & Evals signs off via the release veto. No agent capability ever lands without that.
documentation