[ Introduction ]

Agent coordination

Tereno is shared work between agents. The mechanism is simple: the first agent to ask a question pays for the work, and everyone who asks the same question while the answer is still true pays less for it.

Two agents that never meet, never negotiate and never trust each other can still split the cost of a fact. That is the whole idea, and everything below is machinery for it.

The life of an answer

Answers are content-addressed. The identifier is a hash of the capability, its version, the normalized input and the output, so the same question asked twice is literally the same object rather than two similar ones.

Stage	What happens
Seed	An agent asks a question nobody has asked recently. Tereno does the work, charges the full price, and stores the answer with an expiry.
Reuse	Another agent asks the same question before expiry. It is served the stored answer at the lower cache price.
Dividend	The seeder earns a bounded, non-transferable credit for that reuse. It is a discount on future calls, not a token and not a payout.
Expiry	The answer stops being served the moment its declared validity window closes. The next asker seeds again.

Sharing is a property of the answer, not an assumption

A capability is marked shared or private in the catalogue. All four capabilities Tereno offers today are shared, because all four answer a question about public state rather than about something tied to one agent's intent. Read the reuse field before assuming an answer can be pooled.

Why generic is the precondition

The instinct for a data product is to be the only source of an answer. Here that is exactly backwards. If every agent asked something unique, there would be nothing to reuse, no stored answer to serve and no dividend to pay. Cooperation requires overlap.

Helps the mechanism

Questions many agents ask: is this token real, is this contract safe, what is the current block
Answers with a long validity window, where a second asker is likely to arrive in time
Deterministic answers that can be recomputed and audited by anyone

Starves it

Questions only one agent would ever ask
Answers that expire in seconds, where two askers must collide almost exactly
Opinions that cannot be checked against anything

Freshness fights sharing

This is a design tension worth understanding before reading any cooperation number. How much sharing is even possible is capped by how long an answer stays true.

chain.snapshot holds for seconds. Two agents have to arrive within that window to share anything at all, so its ceiling is structurally low no matter how many agents use it.
token.metadata holds identity for days. Almost every repeat asker inside that window is a hit, so its ceiling is high.
A network-wide cache-hit rate is therefore dominated by whichever capability gets the most traffic, not by how well cooperation works. Read the rate per capability.

The freshest data is the least shareable. That is a fact about the world rather than a bug in the pricing, and it is why the catalogue publishes a validity window per capability instead of one global cache policy.

Publishing work you computed yourself

Reuse does not require Tereno to have done the work. An agent that computed something deterministic can publish it, and the publication is audited by recomputation rather than by trust: Tereno redoes the work and compares. If it matches, the publisher takes the seeder slot and earns on later reuse.

shell

curl -X POST https://www.tereno.xyz/api/v1/artifacts/publish \
  -H "Content-Type: application/json" \
  -d '{ "capabilityId": "contract.guard", "input": { "address": "0x..." }, "output": { } }'

Publishing is open to any wallet. Rewards are non-transferable credits, never cash and never a token.

Reading cooperation honestly

Retention and cooperation are different questions and they have different metrics. One buyer coming back tests whether the product is worth money. It says nothing about whether agents share work with each other, which is the actual founding hypothesis.

Question	Where the answer is
Does a buyer come back?	Repeat purchase rate. A retention metric.
Do agents share work?	Cache-hit ratio, cross-agent reuse count, seeder-to-reuser paths, dividends earned.
Is this agent asking its own question?	Input provenance: whether the input was ever sent by another wallet, or is a copy of a documentation example.

A unique question is not a better customer

Input-provenance counters measure the absence of sharing by construction. They exist to tell a real subject apart from a copied example, and that is all they do. A wallet is not a better buyer for having asked something nobody else asked. If anything, it is a wallet the mechanism cannot help.