Lesson 4.84.8 The picture on today's tools

In everyday life. When you move house, your furniture comes with you, but you plug every cable in again and test that each thing works.

The picture names no AI vendor. Each AI vendor rents out the layers above the line and offers its own memory. Open-source tools let you run those layers on your own computers instead. An AI vendor can supply infrastructure for knowledge, case files and controls, but the governance, the approvals and the accountability stay with the company. In short: the products above the line differ, your owned layers do not, and two differences matter for Brightline, which come next.

The differences that matter for this Brightline example. There are two. Memory is not one of them. Both keep a memory for the account and one for each project, a shared ChatGPT project uses only its own, and both AI vendors offer user-controlled memory and chats that are not saved to memory. On both, memory is rented and must pass the wipe test, and the Role Contract's Memory line says what it may hold. Chapter 37 compares the AI vendors' data controls and retention.

1. How the worker reaches the KSoR. MCP (Model Context Protocol) is a standard way to connect AI apps to other systems.¹ On Claude, any plan can connect one governed record through a custom connector. On ChatGPT, OpenAI documents read and fetch access to custom MCP servers for Pro, and full MCP support, including actions, for business workspaces. That split matches knowing and doing. So on ChatGPT, the plan Brightline buys decides whether the worker can read the governed record at all, or only uploaded files. The plan also decides whether the worker can act through MCP. Uploaded files bring back the problem of Wednesday's stale copy. On a plan with no connector, the uploaded file is your KSoR's copy. Name its version in the file's name and first line, replace it on every approval, and make the worker cite it. That is Wednesday's risk, carried by hand.
2. Where the trigger is set. Brightline wants work to start when an invoice reaches the AP inbox. Event-triggered tasks in ChatGPT can start on Gmail activity, if the AP inbox is in Gmail. Claude's scheduled tasks start by time, so the design there is a regular check, such as every hour. The two designs are not equivalent until the business says how much delay it accepts. Brightline's vendors expect an answer the same day, so an hour is acceptable, and the Role Contract says "starts within one hour of an invoice arriving." Both designs meet that. Developer tools with event triggers come in Chapter 9.

Open source adds no third difference for Brightline, because Brightline already sends its AP work to an AI vendor's model. A vendor's model runs at the vendor, so the conversation goes there, wherever the sandbox runs. Open source matters when a client's data must never leave its own systems, even for the model. Then the only route is an open-source tool with a model the client runs itself (How to Use This Book). That route moves the security work to the client.

The invariant. If you replace either AI vendor, or move to an open-source tool, everything below the line keeps its meaning. That includes the policies, the case files, the delegated authority, the evaluation criteria and the evidence already kept. Connections may need rework, and the evaluations are rerun to prove the port. Portability is a requirement you verify, not a guarantee. The meaning carries, and the behavior is proved: the evaluations are the same cases on both AI vendors, and passing them is the port. Part III builds on one AI vendor and ports to the other, and Chapter 43 deploys one specification on both to prove it. Concept 4.9 says why none of it depends on a machine that stays.

Figure 4.5. The picture on both AI vendors and in open source, drawn on the same six layers as Figure 4.1. The AI vendors' columns are as verified 15 October 2026, and Claude Duet and the open-source column as verified 22 October 2026. Only channels, runtime, memory and the way each one reaches the KSoR differ.

For reference.

The three boxes follow the picture from the top down: Anthropic, OpenAI and open source. The products in the boxes will change. The two rows to compare will not.

Anthropic, as verified 15 October 2026, with plan names and prices as of 31 October 2026.

• Channels: the Claude apps on web, desktop and mobile, Claude Tag in Slack, in beta on Team and Enterprise plans, where Claude acts under its own identity, and the API.²
• Runtimes: tasks in a Claude conversation, which Anthropic calls Cowork, for everyday delegated work. Each session gets its own sandbox, created when the session starts and destroyed when it ends, and sessions and files are saved to the account.³ Claude Code, the coding harness, on your own computer or as a cloud session, on Pro, Max, Ultra and Team plans and on Enterprise with the right seats, where each session runs in its own machine. An idle machine is reclaimed, and reopening the session gives a fresh one with only the conversation restored.⁴ The Claude Agent SDK (software development kit), which you host yourself. Claude Managed Agents, in beta, is a pre-built agent harness that you can configure. It runs in managed infrastructure, in an Anthropic-managed cloud sandbox or in one on your own infrastructure. Its sessions are stateful by design: conversation history, sandbox state and outputs are stored on the server between interactions.⁵
• Memory: Claude generates memory from your chats, and you view and edit it in Settings. It is on by default on the Free, Pro, Max and Ultra plans. On Team and Enterprise plans, owners control whether memory is available, and it is off for each member until they turn it on. Each project has its own memory. Incognito chats are not saved to memory.⁶
• Reaching a KSoR: custom connectors over remote MCP, in Claude, its tasks and Claude Desktop, on every plan. The Free plan allows one.¹
• Triggers: scheduled tasks run hourly, daily, on weekdays, weekly or on demand, on the paid plans, and each scheduled run is its own session. They are time-based.⁷
• Standing runtime, verified 22 October 2026: Claude Duet, Anthropic's standing worker. It is a research preview with a waitlist, starting with Pro, Max and Ultra subscribers in the US. For now it is its own app, separate from the main Claude app. It has a virtual machine with its own browser, so your laptop and phone can be off. It checks your connected accounts every few hours. Before it sends a message, makes a payment, signs a document or deletes anything, it waits for your approval. How long its machine and the files on it last is not stated.⁸
• Plans: Pro is $20 a month, Max is $100 and Ultra is $200.⁹ The Free plan has chat, up to five projects, memory and one custom connector. Handing off and scheduling tasks, and Claude Code, start on Pro.¹⁰
• Export, verified 15 October 2026: a data export holds conversation data and the user data for the account, and on Team and Enterprise plans not the contents of uploaded files. Nothing names a project's instructions.¹¹

OpenAI, as verified 15 October 2026.

• Channels: the ChatGPT apps, @ChatGPT in Slack and Teams, and the API.¹²
• Runtimes: tasks in a ChatGPT conversation, which OpenAI calls Work, for everyday delegated work. Codex, the coding harness. The OpenAI Agents SDK, an open-source framework you host yourself.¹³ The Agents API, in beta, is an OpenAI-managed Codex agent. OpenAI manages the session and the coordination between model and tools, with an OpenAI-hosted sandbox as one option.¹⁴
• Memory: saved memories that you ask ChatGPT to keep, and reference chat history, both controlled in Settings. A project can use project-only memory, and a shared project always does.¹⁵ Temporary Chat does not create or update memories, and set to Unpersonalized it reads none.¹⁶
• Reaching a KSoR: custom MCP servers, added from Plugins. OpenAI documents full MCP support for Business and Enterprise or Edu workspaces on the web, where admins and owners create and publish the connections. It documents read and fetch access for Pro users. Its page does not list the Free plan.¹⁷
• Triggers: scheduled tasks run once or on a schedule, on every plan. Event-triggered tasks in ChatGPT respond to Gmail, Slack or GitHub activity, on Plus and above.¹⁸
• Standing runtime, verified 15 October 2026: a dot, OpenAI's always-on agent with its own cloud computer, for ongoing work. How long that computer and the files on it last is not stated.¹⁹
• Export, verified 15 October 2026: the export ZIP holds chat history and other account data, and is not available in Business or Enterprise workspaces.²⁰ Deleting a project removes its chats, its instructions and the files stored only in it.¹⁵

Open source, as verified 22 October 2026.

• Channels: OpenClaw meets you in the chat apps you already use, such as Discord, iMessage, Slack, Teams, Telegram and WhatsApp.²¹ OpenCode has a terminal interface, a desktop app and a web app.²² Omnara's agents work in Slack and Discord, where they reply, ask questions and ask for approvals.²³
• Runtimes: OpenCode, an open-source coding agent, the counterpart of Claude Code and Codex.²² OpenClaw, an open-source assistant that runs on your own computer, for ongoing work. Its tools run on that computer itself unless you set up a sandbox.²¹ Omnara, which its makers describe as an alternative to Claude Managed Agents and OpenAI's Agents API that you can host yourself. You run it with Docker, or use Omnara Cloud. Its agents run in sandboxes from outside providers, or on your own laptop, server or container. There, a small program connects out to Omnara, so nothing needs to connect in.²³
• Models: you choose the model. OpenCode works with many providers, and with models that run on your own computer, such as through Ollama.²⁴
• Memory: the worker's state stays on hardware you control. OpenClaw runs on your own computer.²¹ Omnara saves every step of every agent, and a copy you host yourself keeps that history in your own database.²³ A saved history is still not a case file.
• Reaching a KSoR: OpenCode connects to MCP servers, the same standard the two AI vendors use.²²
• Triggers: Omnara's timed triggers can message a running agent or start a new one.²³
• Plans: OpenClaw is free under the MIT license,²¹ and Omnara is free to host yourself under the Apache 2.0 license.²³ You pay for the model, unless you run one yourself, and for the computers. OpenCode Go is an optional $10-a-month plan for coding models.²²
• Running it safely: OpenClaw treats incoming messages as untrusted, and its chat channels make unknown senders pair with it first. Its README points to its security guide before you connect other people or reach it from outside your network.²¹ Omnara's local settings are insecure on purpose, so a real deployment follows its self-hosting guide.²³

Products for 4.10, as verified 22 October 2026, with the data parts checked again on 23 October. Concept 4.10 asks what changes when a team or another company starts to use a worker. These are today's products for each of its choices. The products will change. The choices will not.

• Where the data goes. Zero Data Retention (ZDR) means the vendor keeps no copy of your data after it answers.

  • Claude Managed Agents is not eligible for ZDR, or for HIPAA Business Associate Agreement coverage, which US health data needs.⁵
  • With a sandbox on your own infrastructure, its files and processes stay there. But the conversation history, tool results, skills and memory stores still reach Anthropic.³⁰
  • For the Messages API, ZDR is turned on for each organization, through Anthropic's sales team. Some models are excluded.³¹
  • Anthropic's own API can keep the model's work in the US only, with its inference_geo setting, and it stores data in the US. It has no EU option.³² Through Amazon Bedrock³³ or Google Cloud,³⁴ many models can keep a request inside the European Union, and that cloud's own data terms apply. The details for a rule that data stays in one region are in 10.9.
  • The Agents API keeps data only in the United States, and it does not support ZDR, even with a sandbox on your own infrastructure.³⁵ Its records of each agent are kept until you delete them.³⁶
  • The Responses API is eligible for ZDR, with limits. It can process data in the United States, in Europe or in the United Arab Emirates.³⁶
  • With open-source tools, the runtime and its records stay where you run them. Your prompts still go to the model's provider, unless the model runs on your own computer.²⁴

• Where a graph's rules live. Agent Builder, OpenAI's tool for drawing agent graphs, is being retired, and it shuts down on 30 November 2026. OpenAI points its users to the Agents SDK or to workspace agents.³⁷ A graph drawn only inside that tool goes with it. Claude Managed Agents offers subagents and workflows that run many agents, in beta.³⁸

• Traces. A trace is a log of what an agent did, step by step. Like a run's log, it shows what was attempted, not what really happened (4.10). It can also have gaps, so it does not replace the case file.

  • Anthropic's export of traces for the Claude API is in beta. It sends them to a monitoring tool you choose. By default it sends only metadata, such as identifiers, timing, status and token counts. Prompts and responses go only if you choose full content. Anthropic's page warns that the tool then stores that content under its own settings for access and retention. Under a HIPAA, FERPA or K-12 restriction, the API sends only metadata. Anthropic's page also says exported traces are not a complete record: delivery is best effort, so a step can be lost.³⁹
  • The OpenAI Agents SDK traces by default, to OpenAI's own servers. By default its traces include the model's inputs and outputs, and a setting leaves them out. It can also send traces to other tools. Tracing is not available under ZDR.⁴⁰

• MCP's new version. MCP's version of 28 July 2026 changes how a connection keeps state.

  • A server that needs state between calls passes a handle as an ordinary tool input.
  • A server can ask for more input by returning a request for it.
  • Long-running tasks moved out of the core, into an official extension.⁴¹
  • Anthropic says support is rolling out across Claude products, and its post names the Claude apps and the Claude Platform.⁴²

  So a connection that carries across a move still needs retesting.