Why must an AI agent never invent a number, and how do you stop it?

Because a made-up figure in an email, an order or a CRM field looks exactly like a real one, and by the time anyone checks, it has been acted on. The fix is to make every number an agent writes traceable to something it read in that same run, and to refuse the write when it is not.

What "grounded" means in practice

A grounded write is one where every figure in the outgoing arguments (a count, a quantity, an amount, a date) can be matched to a value the agent obtained during the run: from a system it queried, from the input it was given, or from arithmetic over those. tacitrun checks this at the moment of the write. A figure with no source is not a style problem; the write is held, the person sees which value has no origin, and the agent’s reasoning is there to read.

Identifiers get the same treatment in a softer form: an order number or record id the run never saw is flagged for the reviewer rather than blocked outright, because ids are often typed in by the person who started the run.

Why this belongs in the runtime, not the prompt

Telling a model "do not make things up" reduces the rate; it does not make it zero. A rule enforced where the write happens makes it zero for the writes that matter, and it is auditable: the trace shows the check, the values and the verdict for every write, every time.

The same principle applies to the agent’s rules. When a rule says counts must be fetched live unless the input supplies them, the runtime guard that enforces it is shown the run input, so the exception is judged on evidence rather than on the agent’s say-so.

The terms, as the product defines them

Trace / Span

The detailed record of what happened inside one run.

Every invocation produces a trace made of spans — each LLM call, tool call, eval, retrieval, and learning step — with cost, latency, and status. Visible in Observability, the Console, and the trace view.

Needs your OK

Where your agents ask you to approve, change, or stop something.

Your inbox of decisions an agent has paused on before acting — each shows what it wants to do and why, in plain words. You can approve it, use your own choice instead, or stop it. Risky actions (like sending a message or moving money) always wait here for a human.

Related

See it on one of your own processes. Free for the whole product for a trial period, no card needed to start, every write held for your approval.