Lesson 22.1922.19 How today's tools realize it

Reference. How today's products realize the chapter's ideas (22.19), then Check yourself. Beyond this chapter, the word list and the field guide (F0 to F7) are on the companion page, Chapter 22 reference (in the lab, docs/chapter-22/reference.md).

Reference. Skim it now, and come back when you choose products.

What you will learn. How three of today's products, one from each kind of vendor, realize this chapter's ideas, how sure each claim is, and the differences that should change your choice.

Why it matters. The products look alike, but they differ in who holds each person's token, what runs before a tool, and what a run costs. Those differences decide which one fits your product.

How sure is each claim?

Has any of this been tested?

Nothing here has been tested in the lab yet. The ports in the lab show that the same worker can be described on each runtime. They do not prove that the three support it equally. No vendor-neutral standard defines a hosted agent. So this section compares each product with the shape the concepts teach. The products look alike, but that resemblance is not a shared standard.

The short boxes below name the products. The full boxes, with every setting, status and price, are on the companion page, Chapter 22 reference. Product detail changes often, so each carries its date: checked November 5, 2026.

Anthropic: Claude Managed Agents. A hosted harness on the Claude API, in beta. Agents, environments, sessions and events,¹ with scheduled deployments as a fifth resource.² Check on its pages how a schedule treats an hour that the clocks skip or repeat. Agents and environments are files applied with the ant command-line tool.³ Permission policies are always_allow, always_ask and auto,⁴ and networking is limited to an allowed_hosts list.⁵ Web search and web fetch run on its servers, not in the sandbox, under the same list.⁶ Since October 7, 2026,⁷ web fetch also opens only addresses already seen in the session, such as one in a person's message. Vaults hold each person's mcp_oauth credential and add it on the way out.⁸ It makes each connector call from its own servers, and a session whose connector's host is not on the list is rejected when it is made.⁹ Each session can carry a budget in whole cents, set when it is made.¹⁰ No free plan: tokens and web searches, plus a charge per session-hour.¹¹

OpenAI: the Agents API and sandbox agents. A hosted harness built on Codex,¹² in beta.¹³ It can load a whole plugin bundle,¹⁴ holds vault credentials when the platform makes the connector call,¹⁵ and also offers sandbox agents, where your own program runs the loop with the second SDK of Chapter 21.¹⁶ No free plan.¹²

Open source: Omnara. A self-hosted platform for running agents, under the Apache License 2.0. Skills and connectors are added one by one, and any model in the Responses, Chat Completions or Messages format (the request shapes of OpenAI's and Claude's APIs, Chapter 20) can run it. Security. Its local defaults are deliberately insecure.¹⁷ It can hold a refreshable token as a secret,¹⁸ but one for each person is not yet tested, so the lab's port uses a short-lived token in a local file. Free to run yourself: you own uptime, patching and security.¹⁷

Which differences change a decision?

What must stay true on any platform?

A hosted worker acts for one person in one company per session, with that person's own token, which the worker never sees. The platform runs the loop, but your application decides which person, which company, which tools, where it may reach and how much it may spend, and it checks every result.

Exam note. The Claude developer certification exam (CCDV-F) is getting an updated guide. Until Anthropic publishes it, this book maps to the guide it expects. There, building hosted agents falls under the objectives for building agents and for agent patterns, and their identity, secrets and keys under the objective of that name. The exam examines Claude, not OpenAI or Omnara, and treats research-preview features as out of scope.