How is an Agent OS different from an LLM provider?

An LLM provider (OpenAI, Anthropic) gives you a model. An Agent OS gives you the infrastructure agents need to actually do things on behalf of users — identity, payments, memory, dispatch, trust. The model is one input; the OS is the runtime.

Does my agent need an OS, or can I just call APIs directly?

For a single product with one agent, direct API calls work. As soon as you have multiple products, multiple users with different entitlements, payments to settle, or AI-driven actions to insure, you start writing an OS. The question is whether to use a shared one or rebuild it.

Is GeraOS open source?

The agent manifest spec and skill schema are public and open. The platform implementation is proprietary; the API is documented and any agent can call it. Pricing is free tier + Studio £29/month.

How does an Agent OS handle trust?

Through signed action receipts. Every dispatched action emits an HMAC-SHA256-signed receipt that downstream consumers can verify. GeraOS receipts are insurable via Gera Action Warranty for £2-£14 per action. Tampering is detectable; disputes are auditable.

Can I run an Agent OS locally for development?

Yes. GeraOS has a local SDK that emulates dispatch, identity, and trust. Mock payments are supported. Production uses the hosted GeraOS via API key.

Definition · 7 min read

What is an Agent Operating System? (2026 Definition)

An Agent OS is the platform layer between AI agents and the products they need to call. Identity. Memory. Payments. Dispatch. Trust receipts. Five layers, one runtime.

Published 9 May 2026 · GeraOS editorial

The short definition

An Agent Operating System is to AI agents what Linux is to desktop programs: the shared infrastructure that lets the things on top do work without each one rebuilding the same plumbing.

You can build an AI app without an Agent OS — many do — but as soon as your agent has to identify users, remember context across sessions, pay vendors, dispatch to multiple back-ends, and prove its actions to third parties, you start writing an OS. The question is whether to share one or roll your own.

The 5 layers

1. Identity

A user-account abstraction that persists across products. Without it, every product has its own login; agents that work across products have to juggle five sets of credentials. GeraOS uses one identity for all 37 Gera Systems products.

2. Memory

A scoped context vault per user. Stores preferences, past actions, conversation summaries. Agents read from it without the user re-explaining context every session. Privacy boundary: each product gets a scoped read; the user controls what's shared between products.

3. Payments

Routing across global rails. An agent doesn't need to know that the user is in Kenya and should pay via M-Pesa, or in India and should pay via UPI — the OS handles routing. GeraOS uses Stripe Connect Express + country-aware rails (Idram, M-Pesa, UPI, PIX, JazzCash, bKash).

4. Dispatch

The agent-to-skill routing layer. When an agent calls "book a plumber in Lagos for tomorrow morning under £50", dispatch resolves that to the right product (GeraHome), the right pro, the right payment rail, and returns a confirmation. Skills are published to a marketplace; the OS picks one.

5. Trust

Every dispatched action emits a signed receipt. Format:

{
  "signature": "hmac_sha256(action_receipt)",
  "source": "gera-home.booking.confirmed",
  "user_id": "u_8f2c1d",
  "amount": 4500,
  "currency": "GBP",
  "timestamp": "2026-05-09T11:42:18Z",
  "resolution": "in-person-confirmed",
  "warranty_id": "warr_b8c4e2"
}

Receipts are insurable, verifiable, and auditable. Gera Action Warranty insures them at £2/£6/£14 tiers. Gera Services Passport ties the seller's verified identity to every receipt they sign.

Why this matters in 2026

The number of agent-driven transactions is growing roughly 3× per year in 2026. Every transaction needs identity (who's paying), routing (which rail), and trust (is the seller real, what happens if the action fails). Without an OS, every developer rebuilds these. With one, every developer ships faster and the ecosystem composes.

FAQ

How is an Agent OS different from an LLM provider?: An LLM provider (OpenAI, Anthropic) gives you a model. An Agent OS gives you the infrastructure agents need to actually do things on behalf of users — identity, payments, memory, dispatch, trust. The model is one input; the OS is the runtime.
Does my agent need an OS, or can I just call APIs directly?: For a single product with one agent, direct API calls work. As soon as you have multiple products, multiple users with different entitlements, payments to settle, or AI-driven actions to insure, you start writing an OS. The question is whether to use a shared one or rebuild it.
Is GeraOS open source?: The agent manifest spec and skill schema are public and open. The platform implementation is proprietary; the API is documented and any agent can call it. Pricing is free tier + Studio £29/month.
How does an Agent OS handle trust?: Through signed action receipts. Every dispatched action emits an HMAC-SHA256-signed receipt that downstream consumers can verify. GeraOS receipts are insurable via Gera Action Warranty for £2-£14 per action. Tampering is detectable; disputes are auditable.
Can I run an Agent OS locally for development?: Yes. GeraOS has a local SDK that emulates dispatch, identity, and trust. Mock payments are supported. Production uses the hosted GeraOS via API key.

Build on GeraOS

Free tier + Studio £29/month. Identity, memory, payments, dispatch, trust receipts — five layers, one API.

See pricing →

Related Gera Systems products: Gera Action Warranty (insures dispatch outcomes), Services Passport (verified seller identity), Gera Skills (the marketplace dispatch routes to).