Inside one company, an agent has a chance. Someone can scope what it can access, review what it did and shut it down when it gets something wrong. That is hard, but it is containable.
Across a company boundary, the problem changes. Your agent asks a partner's agent to confirm a fee. Their agent checks the terms it can see and sends back an answer. Your agent accepts it, and both sides record the result.
There is only one problem: neither agent saw the amendment both companies signed last quarter. The agents communicated perfectly. The answer was still wrong.
A protocol can carry the message but not the agreement behind it
A lot of work is going into protocols that let agents discover each other, exchange information and coordinate tasks. That matters. But a protocol can tell one agent how to talk to another. It cannot tell either agent what the two companies actually agreed.
That relationship is specific:
- the fee schedule
- the service levels each side committed to
- the liability terms
- the amendment negotiated six months later
- the exception that applies only to this partner
- the version that was in force when the transaction happened
Generic protocols are generic by design. Commercial relationships are not, and once agents begin making decisions across company boundaries, that difference becomes operational.
People carried the relationship context
Humans had a workaround. Someone in payments knew someone at the processor. Someone in finance had the latest amendment. Someone remembered that this partner had different terms. When the records did not line up, somebody picked up the phone. It was inefficient, but there was a social layer around the systems that could recover missing context.
Agents do not have that. They do not remember the conversation from two years ago unless it was captured somewhere they can access. They do not know an agreement changed unless the new version is part of their context. And they should not be expected to infer commercial terms from the behavior of the systems around them. An agent without the agreement has plenty of intelligence and nothing that defines the right answer.
The agreement has to become part of the infrastructure
For agents to work safely across companies, the relationship itself has to become something a system can use. That means more than the contract PDF. It means the current terms and their amendments, the operating rules derived from them, which version applies, which system behavior meets the agreement and which does not, and which exceptions require a person.
That context has to travel with the work. Otherwise two highly capable agents can exchange perfectly structured messages and still reach the wrong conclusion.
Cordant holds what the relationship means
Cordant is built to hold what two parties actually agreed, amendments included, and connect those terms to what each side actually did. It is built so that when an agent needs to know whether a fee was correct, the answer can be checked against the agreement that governs the relationship rather than inferred from whichever system happens to be available. Where the two sides differ, Cordant shows where, which term applies and what evidence supports the answer.
Agents need more than a way to talk to each other. They need a shared reference for what they are talking about.
List the agreements your partners' agents would need
Assume your major counterparties turn on agents next quarter. Which agreements would those agents need to know about, and which amendments, fee schedules, versions and exceptions? Make the list. It is also the list of places where two agents could communicate flawlessly, agree with each other and still be wrong.
The next problem in agent infrastructure is not getting agents to talk. It is making sure they are talking about the same agreement.
