Towards Agentic Determinism

TL;DR

Software engineering was never deterministic. Software delivery is. No two engineers write the same code. Production quality comes from the QA gate, not the person.
Digital Workers follow the same law. They are probabilistic by nature. That is fine, as long as every output passes a QA test that verifies it against explicit criteria before it touches an operational system.
The QA test is the deployment decision. If you can write the test, the Digital Worker can act autonomously. If you cannot, it should inform decisions, not execute them.
No test, no write access. Workers without a verifiable QA gate still create value as decision support or copilots. They produce the inputs; humans make the call.

Our Argument:

Engineers were always probabilistic

Give the same ticket to two engineers and you get two different implementations. Different structure, different naming, different edge cases. Software production has always been probabilistic at the point of creation.

Nobody panicked about this. The industry solved it with a QA function: tests that code must pass before it reaches production. Determinism was never a property of the engineer. It was a property of the gate.

The same law governs Digital Workers

Digital Workers deployed across business areas are probabilistic by nature. Run the same task twice and the reasoning path will differ. This is not a flaw to engineer away. It is the same property that made human engineers productive, and it demands the same answer.

Every Digital Worker output should be verified with a specific QA test before it writes into an operational system. Not a vibe check, not a spot review. A defined test that confirms the output satisfies explicit criteria: the reconciliation balances, the submission matches the source data, the contract terms fall inside policy.

Design the test well and the probabilistic nature of the worker stops mattering. What reaches production is deterministic, because nothing reaches production without passing the gate.

The market is learning this the hard way. Forrester and Anaconda found that 88% of agent pilots never reach production, with evaluation gaps the single most cited blocker, and 70% of leaders name non-deterministic outputs as the top barrier to production readiness. These teams are stuck because they are trying to make the worker deterministic instead of making the gate deterministic.

No test? Then no write access

The honest question for every deployment is: can we write the QA test?

If yes, the Digital Worker can operate autonomously inside operational systems, gated by verification.

If no, the worker should not write into operational systems at all. It still earns its keep as decision support: producing reports, analysis, and recommendations that feed human decision-making. Or it works as a copilot, with a human validating each action. The value is real. The write access is not.

That single question, test or no test, is a cleaner deployment framework than any maturity model.

At causaLens, this is how we build: every Digital Worker ships with the verification layer that makes its outputs deterministic. Reliability is not a property of the model. It is a property of the gate.