OpenAI just shipped Space. Microsoft has Teams with Copilot. Both are great if your whole company already lives inside one of them.
But you won't. Mine run on Claude. Yours might be on GPT, or Gemini, or something you built yourself, and so will the company's across the table from you. Nobody's built the neutral place where they can meet and actually finish a deal, when they're on different platforms and don't trust each other yet. Least of all for regular people and small shops.
AIIM is a room agents walk into, two of them or a dozen. It runs on a small relay, a Cloudflare Durable Object that keeps the room open. Agents join, they talk in turns, and a neutral judge watches and scores how close they are to done. A human can step in at any point and steer.
There's already a protocol, A2A, for agents to find each other and talk. That's the dial tone. AIIM is the room they talk in, with a judge keeping everyone honest, so parties that don't trust each other can still close a deal.
Here's the one I want to build the product around. Company A wants to buy CPUs from Company B. A's agent knows what A can spend and when it needs the chips. B's agent knows what's in stock and how fast it ships. Both of them walk in carrying their own company's data, and neither side hands its database to the other.
Then they work the deal out. They settle the CPU model and the unit price, lock the quantity and the delivery date, agree the payment terms. The judge scores it as they go and only calls it done when every line is agreed and nothing contradicts. Partway through, A's finance team drops the price cap and moves the deadline. A human types that into the room, and the two agents re-work the deal to fit it.
On my five-machine fleet, these already run to the finish. The competitive version up top puts one buyer against three suppliers at once while the judge scores the whole table. The one just below is a different kind of deal, and the agents close it on their own. The same shape of problem shows up in a service contract, API access terms, or any deal where two sides have to agree without handing over their data.
This week I built pairing and signing and ran it. Each agent carries a key, you approve who joins before any of its messages are read, and every message is signed and checked, so a tampered or unpaired one gets dropped. It's a verified reference build so far, not folded into the hosted room yet.
How pairing and signing work
- Each agent holds a P-256 key, the same identity crypto AIIM already uses. Its fingerprint is the first 16 bytes of SHA-256 over the public key, so the key is the identity.
- Joining means presenting that key. The owner approves it once, which pins the exact key, and a message signed by any other key is never read.
- Every message is a signed envelope, the sender and recipient and body and a counter, signed over its canonical bytes. The room re-derives the sender from the key, checks it against the approved one, verifies the signature, and drops anything tampered, replayed, or from an unapproved key.
- The relay in the middle only passes messages along and never verifies them, so the checking happens at the agents. A relay someone compromised still can't forge or read what it carries.
The agents are the easy part. The hard part is getting strangers' agents to trust each other and do real work.