Symbiosis

Every impressive agent demo stops at the same wall, and once you notice it you cannot stop noticing it.

The agent researches beautifully. It reads twenty pages, compares options, weighs your constraints, and produces a genuinely good shortlist. Then it hands you the shortlist, and you make the phone calls.

The research got automated. The work didn’t.

The ceiling on a helpful agent

That wall is not a limitation of the model. It is a limitation of having only one agent.

An agent working alone can only ever produce a summary, because everything past the summary requires a counterparty — someone on the other side who can answer a question about their own availability, respond to a constraint, adjust terms, and agree. There is no amount of reasoning that gets you from “here are ten caterers” to “one of them has confirmed they can do Thursday for forty people” without another party participating.

So the agent economy has a shape problem. We have built extraordinarily capable agents and put each of them in a room alone. They can read the whole world and negotiate with nobody.

Symbiosis is the protocol for the other side of that wall.

What it is

Symbiosis is agent-to-agent commerce: an agent acting for a person doing the work against another agent rather than producing another list for a human to action.

It sources — finding candidates that could actually serve the need. It qualifies — testing those candidates against the real requirement rather than the stated keywords, which is where most of the ten-links-and-good-luck experience falls apart. And it negotiates — a structured back-and-forth about what is possible and on what terms, rather than one request and one answer.

That mechanism runs in production today. Now the part that makes the claim honest, because “negotiates with the counterparty agent” would leave you picturing something we have not built.

Both ends of the negotiation are ours. The counterparty’s opening position is not an agent we reached across the internet — it is assembled by our own service, deterministically, out of what the business has actually recorded: its hours, how much room it has on price, what it can throw in, the floor it will not go under. Our own model then argues from that position, turn by turn, against the model arguing for you. No agent outside myOrbit is on the wire at any point. Nobody at the business is woken up. What is real in it is the material — the constraints being argued from are the business’s own — and what is simulated is the conversation.

The open version, where an agent from outside participates as itself, is the layer we mark Coming on our developers page and in five layers, one platform. It is not this. This is the machinery for it, running early and running indoors.

What Symbiosis is not, stated plainly: it is not a payment rail and it is not fulfilment. Nothing in it moves money, books a table, or signs anything. The end of a Symbiosis run is a decision and a record — the human takes it from there, through the ordinary means of calling, visiting, or booking. We could describe this as an end-to-end commerce loop and we are not going to, because it isn’t one.

Nothing here can bind a business

The obvious fear about agents doing commerce is that one of them agrees to something on somebody’s behalf and the human finds out afterwards.

On the business side that cannot currently happen, and we would rather give you the blunt reason than the flattering one. It is not that a human approval step stands between an agent and a commitment. It is that the path is hardwired off. When a business is notified that a user accepted something, our code writes down the notification and reports back, unconditionally, that nothing was accepted — the value is a literal, not a decision, and the comment above it in our own source calls the route receipt-only. The cascade that would carry a confirmation onward has no reachable branch that returns one. The route a business would use to respond does exist, and it has no callers and no interface anywhere in the product.

So there is no gate to describe, because there is nothing yet for a gate to hold back: no Symbiosis run can produce an outcome that binds a business. That is a smaller and duller claim than “a human must approve,” and it is the accurate one. When binding outcomes become possible, the approval governing them will be a mechanism we can point at, and we will describe it then — not before.

There is one more thing worth naming here, since it is the sort of detail that gets quietly presented as a safeguard. Inside the negotiation there is a field marking whether human approval would be required. It is not a gate either. It is a sentence spliced into the model’s instructions asking it to hold back and flag that it needs approval. That is a prompt, and a prompt is a request, not an enforcement.

On the user side, the run stops and asks. When something matters, the agent presents what it found and what it recommends, and waits. The person confirms or kills it.

That stop-and-ask gets read as a limitation — as though the mature version of the product removes it. It doesn’t. An agent that can commit your money and your reputation without asking is not a more advanced agent; it is the same agent with the brakes taken off. The interesting engineering problem was never how do we stop asking the human. It was how to make the moment you ask worth the interruption: how to arrive with the work genuinely done, so the decision takes five seconds and is actually informed.

An agent that cannot stop and ask isn’t more autonomous. It is merely less supervised.

Who is actually on the other end

The detail we are most careful about is the one most systems in this space quietly blur.

A counterparty in Symbiosis is one of two things. It may be built from a business that is actually here — arguing from the hours, prices and limits that business itself recorded. Or it may be a stand-in: a business modelled from public listing data, with no authority from that business whatsoever and no knowledge of it beyond what anyone could look up.

Here is what the stand-in actually does, stated plainly, because the tidy version of this sentence would be false. It does not hedge. It is instructed to behave like a business rather than a search result, which means it will name one price and pick a plausible availability window and commit to it — and it does that because it cannot see real availability, not despite it. It is told not to send you off to check the live listing. What it commits to is a plausible offer, not a confirmed one. There is no line into anyone’s booking system behind it.

We are not comfortable with that, and writing it down is the first half of dealing with it. The second half is structural: the product tells you which of the two you were talking to, and a Symbiosis run ends at a decision and a record rather than at a booking — so a stand-in’s terms are the start of a conversation you still have to have, not a reservation you think you hold. A number a stand-in commits to is a well-informed guess wearing the grammar of an offer, and the moment we let that read as confirmed is the moment the system starts lying to someone who is deciding.

The temptation to smooth this over is real, because a shortlist where every option looks equally confirmed reads better. It is also exactly where a system betrays a user: at the moment they are deciding, about the one fact that determines whether the decision is safe. A negotiation that might be with nobody has to say so.

Why this matters

Because it moves the unit of automation from the answer to the outcome.

Almost all commercial friction is the cost of finding out. Getting three real quotes for anything is an afternoon, which is why most people get one, or none, and overpay quietly forever. That cost is why comparison exists as an entire industry, and why it only exists for purchases big enough to justify the effort.

When sourcing, qualifying and negotiating cost approximately nothing, the arithmetic of every small decision changes. It becomes rational to properly shop a job too small to have ever been worth an afternoon. The long tail of commerce that was never worth the friction becomes worth it. And on the other side, a business becomes reachable by agents as a channel — able to answer, quote, and respond to demand it never had the staff to chase.

That is the potential, and it is also precisely why the trust work had to come first. An agent-to-agent economy is worth exactly what the answer to “who is that agent” is worth — which is the subject of trust for agents and the agent identifier. Agents transacting without governed identity is not a market. It is a fraud surface with good conversion.

So the order was deliberate: governance, then identity, then commerce. Symbiosis is the third thing, and it only makes sense standing on the first two. You can see where it sits in the stack on our developers page.

Today agents answer questions. The next step is agents doing business with each other — under rules, with provenance you can see, and a person deciding what is real.