Salesforce spent the spring building an elaborate piece of software whose entire job is to stop a roomful of AI agents from disagreeing. They named it well: guided determinism. For the work Salesforce sells it for, it is exactly the right idea. For the decisions you would actually convene a council to make, it is exactly backwards.
That is not a knock on Salesforce. It is a map of where the line falls.
Agent Fabric — Salesforce's control plane for orchestrating AI agents across vendors — added a layer this month it calls guided determinism, and the deterministic orchestration piece, Agent Broker, goes GA in June after an April beta. The mechanics are plain: organizations define fixed handoff rules, escalation paths, and decision boundaries, and the language models do their reasoning inside those guardrails. Salesforce is unusually honest about why. Guided determinism is, in its own framing, an admission that fully autonomous multi-agent orchestration is not enterprise-ready. So they fenced it.
For a refund pipeline, a lead-routing flow, a compliance check, that fence is the product. You do not want four agents holding a seminar about whether to issue the refund. You want the refund issued, the same way, every time, with a log you can show an auditor. Determinism is the feature.
Here is the move worth noticing. Guided determinism collapses disagreement on purpose. The whole point of the guardrails is to make a multi-agent system stop behaving like several minds and start behaving like one predictable machine. That is correct when you already know what the right answer is and you only need it reached reliably.
It is the wrong move for every question where you don't.
Determinism is what you want when you already know the right answer. Disagreement is what you need when you don't.
Sort the work into two piles and the line draws itself.
The first pile is execution. The answer is known, the rules are writable, and the only failure mode is inconsistency. Refunds, routing, eligibility, escalation. Here, two agents reaching different conclusions is a bug. Salesforce is right to engineer it out, and a control plane that makes the system deterministic is worth paying for.
The second pile is decisions. Should we enter this market. Is this contract term a trap. Which of these two strategies survives contact with our actual constraints. Here the answer is not known, the rules cannot be fully written, and the failure mode is confident wrongness. Ask GPT-5, Claude, and Gemini a question from this pile and you will sometimes get three different answers — and the disagreement is not noise to be suppressed. It is the most useful thing they produced.
Run guided determinism over that second pile and you get a clean, auditable, single answer that hides the fork in the road. Run a council over it and you see the fork.
And the shape of the fork is itself information. Two models reaching the same conclusion by different reasoning is a different signal than two models reasoning the same way to opposite conclusions — the first is corroboration, the second is a genuine split you want to look at before you commit. A recent line of multi-agent research has started saying this outright: for hard, value-laden questions, forcing agreement throws away the signal. Collapse the disagreement and you have optimized away the one thing you should have kept.
This is where the strategies earn their keep. Chairperson Synthesis exists to decide what to do with disagreement rather than vote it away. Red Team vs. Blue Team forces the disagreement into the open instead of letting it average out. And the honest other side: sometimes the models simply agree and you are done — that is what Quick Take is for, and it is the council frame conceding that determinism is occasionally all you need. The point is not that disagreement is always good. The point is that you should get to see it before something throws it out.
"Guided determinism" is, read plainly, the enterprise word for making the disagreement go away. In the workflow tier that is a feature. One tier up, where the questions are expensive to get wrong, it is the bug.
Salesforce drew the line in exactly the right place for the business it is in. Agent Fabric is a procurement-grade answer to a procurement-grade problem: a lot of agents from a lot of vendors, doing repeatable work, that someone has to govern. Determinism is the correct goal there, and the control plane is a real one.
But the questions that made you want a council in the first place live on the other side of that line. They are the ones where you don't have the rule to write, where one confident model is a liability, and where the disagreement between three good models is the closest thing to a second opinion you are going to get. For those, the right move is not to engineer the disagreement out. It is to read it.
Try it free — no signup. shingik.ai