Software that reads an operation is useful. Software that acts on one is far more useful, and it has to be held to a much higher standard. A wrong answer from a chatbot costs a reread. A wrong action in the physical world costs a missed delivery, an idle line or money committed to the wrong carrier.
Here is how we think about letting AI act on the physical world.
One door for every agent
Juno, Oway's industrial AI, and any MCP-compatible agent a company chooses reach an IOI node the same way: through the Model Context Protocol, with permissions the node's owner sets. Our own AI uses the same door, under the same rules, as every other agent.
Through MCP, an agent works with the same IOI schema operations as any other connected system: reading inventory, checking capacity, quoting, booking, tendering a load or asking a plant to produce. What it is allowed to do with them depends on its scope.
Every agent gets a scope:
- ReadQuery state and use IOI SearchNo changes
- ProposeSuggest actions for a person to acceptWaits for approval
- ActTake defined actions inside its grantApproval for money or plan changes, if set
Every read, proposal and action is logged with its inputs and who approved it.
Oway. Scopes can be narrowed or revoked at any time.
- Read. The agent can query state and use IOI Search.
- Propose. The agent can suggest actions, which wait for a person or another approved system to accept them.
- Act. The agent can take defined actions directly, inside the limits of its grant.
Scopes can be narrowed to specific data, facilities or actions, and revoked at any time. Most companies start an agent on read, review its proposals for a while, and then grant it act on specific, well-understood operations.
People stay in the loop where it counts
Some decisions should always have a person behind them. Commitments that move money or change a plan can be set to require approval, and Juno confirms before anything is charged. A proposed fix to a dock conflict, for example, shows up for the site lead with the reasoning and the alternatives, and goes ahead only when they say so.
Rules that come from the real world
Juno follows rules that come from the physical world and from regulation. It adds a stop only when every existing delivery window still holds. It verifies carrier authority, insurance and ELD connection before freight moves. It respects the weight and space left in the trailer and the hours a driver has left. Every action runs inside the scopes the owner grants.
Treat outside content as data
Industrial AI reads a lot of content it did not write: EDI messages, emails, purchase orders, nightly files, PDFs and messages from partners. Any of them could contain text written to manipulate an AI. We treat all outside content as untrusted data. It can inform a decision. It can never change the rules an agent runs under or the permissions it holds.
A full record
Every decision Juno or another agent makes is logged with its inputs, its outcome and who approved it. When a customer asks why a load moved or a window changed, the answer is on the record along with the data behind it.
Why this matters now
AI is quickly getting good enough to run much of the coordination work people do by hand today. The companies that benefit will be the ones that can let it act, confidently, inside clear limits. IOI and MCP are built to give them exactly that.
For developers, the docs cover MCP, the API and scopes. For security details, see our Security page. In the last note of this series, we step back and look at what all of this means for American industry.
