Tools instead of knowledge
The agent does not know the stock by heart, it queries it. That makes its information as current as the system behind it, instead of as old as its training.
Autonomy
Agents: what they may do and what they need
An agent that acts is something different from an assistant that answers. Where exactly the line runs — and which interfaces make it possible in the first place for a machine to make a request instead of filling in a form.
The difference
Answering is not acting
A chat window that gives information saves a click. An agent that finds the follow-on load, contacts the carrier and watches the deadline saves an entire process — and that is a different order of magnitude.
Several steps
Search, assess, enquire, follow up: a transaction rarely consists of one step. The benefit arises where the chain runs through without a query back.
A traceable trail
Every call, every enquiry sent, every answer is logged. An agent without a log is worthless in a dispute.
The limit
Autonomous is the enquiring, never the awarding
A request is reversible: it can be declined, it lapses, at worst it costs attention. An award is not: it binds, costs money and carries a name.
What the agent may do
Search for candidates, propose follow-on loads, send enquiries with a target price and deadline, collect answers, watch deadlines.
What it may not do
Award orders, commit finally to prices, offer contracts, change master data. These steps carry a signature.
Why this is not a step backwards
The time gained sits almost entirely in the enquiring — a follow-on load is only worth something for hours. The award still has time for a human.
This limit sounds like caution and is a calculation: as soon as two bidding programs compete against each other without a signature, the auction no longer ends where a human would stop. Two agents, one load — played through.
In practice
What this looks like in day-to-day business
The follow-on load after unloading
When a vehicle becomes free, the agent looks for the matching follow-on load nearby and within the time window — and enquires with the most promising partner instead of producing a list nobody works through.
The cascade without waiting time
A refusal or a lapse immediately triggers the next enquiry. No transaction is left lying because someone forgot to follow up.
The assistant alongside
The window remains for questions: why this trailer and not the nearer one? The agent explains its own judgement from the stored quantities.
The standard
One plug instead of twenty special solutions
Building every connection individually was normal for decades and is the reason why integration projects are expensive. With the Model Context Protocol (MCP), a format has established itself that describes which tools a system offers and how to call them — readable by agents, independent of the vendor.
Self-describing
The system tells the agent what it can do. It does not have to be programmed for it beforehand — that saves exactly the work that otherwise makes every connection expensive.
Vendor-independent
The same connection carries different agents. Anyone building on a single platform swaps one dependency for the next.
Connected instead of bypassed
A system without such an interface is not replaced — it is skipped, and the data flows past it.
We stick to it: our logistics glossary, the site search and the free calculators stand ready as an open MCP server — connectable without registration to Claude, ChatGPT or an agent of your own.
In practice
What we do with it
Through this route our portal hangs, among other things, on Transporeon's automatic quoting: key figures on lanes and quoting behaviour come in as a market view that our own history cannot supply.
A market view on your own price
Your own target price can be mirrored against quoting behaviour in the market — as evidence for an adjustment, not as an automatic intervention.
Provenance stays visible
Whatever comes from outside is marked as such and is never written into a decision unchecked. External data is a view, not a verdict.
Confirmation by people
An agent may propose changes to rule sets and carry them out after release — the same rule as with our own parameters.
The actual work
Rights, not technology
Opening an interface is quickly done. The work sits in the question of who may see and trigger what: which data leaves the house, which call changes something, what happens on an error, and how every access is logged. Whoever builds that in afterwards has not built it in.
That is why the same separation applies here as everywhere else: read access generously, changing access narrowly and logged — and everything that becomes binding with a human in front of it.
Direct Inquiry
Bound autonomy cleanly?
We show which steps an agent can take over, where the release stays with the human, and which interfaces are needed for it.