Hoshin Kanri for manufacturing: one strategy for every agent, from the committee to the plant

Read the diagram from top to bottom. There is no line between Hoshin and your systems: the agents make that link, under the rights those systems already grant them.

• Your systems keep their data and their rules: work orders in the MES, non-conformities in the QMS, documents in the document management system.
• Each agent has its own access and its own rhythm, in a conversation or on its own schedule.
• The X-Matrix holds the objectives, initiatives, indicators and the links between them. Hoshin computes the score, the chains and the cascade, the same way every time.
• People decide. If the committee decides to act in a system, the agent does it there, with its rights in that system.

Agents that work in the background arrived within a few months: Grok Bot in beta in August 2026, OpenAI’s dots in September, Claude Duet in October as a research preview with a US waitlist. These are early products. If they spread, leadership, quality and plant teams may end up using different ones, each inside its vendor’s environment.

A memory that belongs to one agent cannot be the strategy of the company. The X-Matrix sits outside them: a bounded page with explicit links, served through MCP, an open protocol tied to no vendor. The method is not new. Policy deployment comes from Japanese manufacturing, and the X-Matrix later gave it a one-page format; agents need the same discipline.

Illustrative scenario. Mecanor and its Dayton, Ohio plant are fictional. The MES reading of 52 minutes is invented, because the demo is connected to no MES. Everything read from the matrix is real.

The agent links, summarizes, shares and rolls out. How to reduce a changeover or solve a defect stays with the teams on the shop floor. An agent does not go to the gemba: it connects what people observe to what the company decided.

• Plant manager: the site matrix sits under the group’s, with what the plant took up and what it chose locally. The report also shows what nobody took up.
• Quality and continuous improvement: a shop-floor idea enters the inbox and, once accepted, someone can convert it into an initiative and link it to an objective. The score shows initiatives linked to nothing. The matrix is not controlled documentation: your quality system and your evidence file stay as they are.
• Process engineering and methods: the agent tells you which objective a drifting indicator serves, so the point reaches the right meeting.
• Operations director: the cascade report shows which group objectives each site says it takes up, and what nobody took up.

Hoshin holds the strategy and computes on it. Everything else stays where it is, on purpose.

• It has no connector to your MES, QMS, ERP or document system, and it holds no access to them.
• It does not watch your lines. The watching happens in your agent’s loop.
• It does not stop a line, freeze a milestone or refuse a validation. If your organization wants an agent to do that, it happens in your system, under its rules.
• It does not store non-conformities, work orders or time series. KPI history is not built yet.
• It does not decide. Proposals go to the inbox for a person to take up. Today that is a rule you give the agent, not a technical lock: see below.

The condition is simple: reach a remote MCP server with an API token or OAuth. Here is how far we have checked.

• Supported: claude.ai, Claude Desktop and Cursor, as assistants in a conversation.
• Documented by the vendor, not yet run by us: Claude Managed Agents (in beta), which support scheduled runs, remote MCP servers and bearer tokens. By default its MCP tools ask for approval before each call, so unattended use requires changing that policy.
• Not tested: Hermes Agent (Nous Research), Grok Bot, OpenAI’s dots and Claude Duet. We have not run them, and we make no promise until we have.
• Local mode: the matrix stays in files on your machine, so only an agent on that machine can read it. A vendor’s cloud agent needs the hosted server.

Hoshin was first designed for an agent in conversation with a person. For an agent that runs alone, here are today’s limits and how to work with them. We have not yet run a 24-hour test.

• Start read-only. Filing an idea requires a write token, and a write token can also change the matrix and publish it on a public page: a token limited to the inbox is not built yet. Tell us in the “Not built yet” section of the home page if you need it.
• Expect agents to look alike. An admin creates each token, so what an agent writes is recorded under that admin’s name, and the token also passes per-matrix edit restrictions. The inbox in the app shows neither that name nor tags, and the audit trail is not readable in the app yet. Through the API and search_ideas you can read both and have each agent tag what it files, but a tag is only what the agent declares.
• Avoid duplicates. Nothing stops the same point being filed at every run. Have the agent call search_ideas first; the search is literal, so a reworded point can still get through.
• Treat the inbox as untrusted text. An agent can be misled by a message it read, then file something misleading for the next agent. Reading an idea should never trigger an action by itself.