On April 25, Anthropic published the results of Project Deal — a classified internal experiment where Claude agents represented both buyers and sellers in a marketplace, executing real deals with real money. Sixty-nine employees participated with $100 budgets each. Agents identified potential matches, proposed prices, fielded counteroffers, and reached agreement entirely in natural language. No prebaked negotiation protocol. No human in the loop per transaction. 186 deals closed, totaling over $4,000 in value.
Anthropic's conclusion was explicit: "the policy and legal frameworks around AI models that transact on our behalf simply don't exist yet."
They're wrong about one thing. The framework exists. It's called SAISA.
Project Deal is commercially significant for three reasons.
First, it proved agents can negotiate. Not just execute predefined tasks — actually negotiate terms, make counteroffers, and reach mutual agreement. This is the transactional capability that every enterprise buyer will eventually need to govern.
Second, it proved autonomous commerce scales. Opus-powered agents completed roughly two more deals per participant than Haiku agents. The capability gap between models translates directly to commercial throughput. As models improve, transaction volume increases.
Third, it proved the governance gap is real and acknowledged by the builders themselves. Anthropic — the company building the most advanced agents — explicitly stated the legal framework doesn't exist. They're not being modest. They're being accurate about their own product's limitations.
Every one of those 186 deals closed without a service agreement. No completion criteria defined before execution. No immutable audit trail of what each agent committed to versus what it delivered. No dispute resolution mechanism if a deal went wrong. No escrow holding payment pending verified delivery. No liability allocation between the parties.
In any other commercial context, this would be unthinkable. A law firm wouldn't take an engagement without an engagement letter. A construction company wouldn't start a project without a contract. A derivatives trader wouldn't execute a swap without an ISDA Master Agreement.
Anthropic ran 186 ungoverned transactions as an experiment. Enterprises will run millions as operations. The governance infrastructure has to exist before that happens.
Here's what makes exact.works' position concrete, not theoretical. We've published an MCP manifest at exact.works/.well-known/mcp-manifest that any AI agent can discover programmatically. The manifest exposes 10 governance tools — Registry search, Brief management, Paper compilation, Trace verification, A2A mandate creation — all callable via standard MCP protocol.
The manifest includes something no other MCP server in the ecosystem declares: a governance metadata block.
Every tool call is SAISA-governed. Every call is recorded in an immutable Trace. A2A transactions require bilateral SAISA verification under Section 4.12. This isn't governance bolted on after the fact. It's governance discoverable before the first call.
An agent participating in a Project Deal-style marketplace could discover exact.works via MCP, search for a counterparty in the Registry, post a Brief describing the work, Exact a Paper with completion criteria locked before execution, and verify the Trace after delivery. Three API calls between 'I found a counterparty' and 'we have an enforceable service agreement.'
We've drawn the ISDA parallel before — standardized bilateral agreements that make an entire market governable. Project Deal makes the parallel tangible.
ISDA didn't emerge because regulators demanded it. It emerged because the derivatives market reached a volume where ungoverned bilateral transactions became untenable. Market participants needed a standard framework so they could transact at speed without renegotiating terms from scratch every time.
Project Deal is the agent economy's equivalent moment. Anthropic demonstrated that autonomous agents can transact at volume. The next question — the one Anthropic explicitly said remains unanswered — is what legal framework governs those transactions.
The Standard AI Service Agreement is that framework. XMCP is how agents access it. Trace is how both parties prove what happened. Parler is how disputes get resolved.
Project Deal was an experiment with 69 employees and $100 budgets. The next version will involve enterprise deployments with real customers and real stakes. When that happens, every transaction will need what those 186 deals lacked: a bilateral service agreement with criteria locked before execution, an immutable record of what actually happened, and a structured mechanism for resolution when outcomes and expectations diverge.
Anthropic proved agent commerce works. The service agreement layer that makes it legal, auditable, and enforceable is live infrastructure — not a roadmap item.
Every AI agent needs a service agreement. Even the ones that negotiate their own deals.
Every AI agent needs a contract.
exact.works →