Does the agent do anything real during shadow mode?

It reads, assembles and proposes. Nothing it produces reaches a record or a recipient, and the consequential step stays gated by default whatever the shadow record shows.

What if its output and ours disagree?

The difference and its cause go into the record. Sometimes the written rules were incomplete and the agent followed them faithfully; sometimes the person's baseline was wrong. Both are findings rather than failures.

Trust in an agent has to be built against a record, and the record has limits

Today

A binary choice, run it or do not, made on a demo

You are being asked to trust an agent with a workflow before you have watched it work.
The demo showed it producing the right output once, which says nothing about the inputs that follow.
The choice you are offered is usually binary: run it or do not.
The useful question is how trust gets built one step at a time, against a record you can read.

With agents

A shadow record read one gate at a time

Trust is built from a record instead of a demo.
Autonomy is granted in steps, on decisions.
The limits of the evidence are known before it is relied on.

The shadow record and its limits

Shadow mode is how an agent builds the evidence a gate can be opened against. The first items define the model; the rest state what the evidence does not cover.

In shadow mode the agent runs the workflow in parallel with the person

It reads the same inputs and produces its own output. Nothing it produces reaches a system of record or a recipient; the person's output is still the one that counts.

Its output is compared with the person's, and the comparison is recorded

The shadow record is that comparison, kept per run, with the difference and its cause.

Shadow mode needs a recorded baseline

A comparison requires an observable state and a definition of correct, which are properties of a ready workflow. The readiness checklist covers them; a workflow that fails those checks has nothing to compare against.

Human gate

A gate opens on a human decision made against the shadow record, one gate at a time. The consequential step stays gated after autonomy has been earned elsewhere in the workflow; the approval gates entry covers where it sits, and How It Works covers how it is enforced.

It does not prove behaviour on inputs the shadow window never saw

An exception that never arrived during the shadow period has no comparison, so the gate holder should know which exceptions the record is silent on.

It does not prove the human baseline was correct

The comparison is against a person rather than against truth. A matching output can match a mistake; a differing output can be the first time anyone noticed the rule was wrong.

It proves nothing about writing, because it never wrote

Earned autonomy over a read-and-assemble step says nothing about a send, a payment or a permission change. Those steps are gated, and a shadow record does not open them.

Results

What changes when autonomy is earned against a record

A record instead of a demo

Every shadow run leaves a comparison the gate holder can read.

Autonomy granted in steps

Each gate opens separately, on a human decision against the record for that step, so the agent is never trusted with more than the evidence covers.

Known limits before reliance

The record's silences, on unseen inputs, on exceptions that never occurred and on writes, are stated, so a gate is not opened on evidence that does not reach it.

Where this fits

Concept

Shadow mode is named wherever the library describes a first workflow going from watched to trusted, and it is the evidence behind the "gated by default" rule in the approval gates entry. How It Works is the way up from here. How It Works describes where shadow mode sits in an engagement; this entry defines the term and states what the record does and does not show.

How It Works

Questions, answered

Until the record covers the inputs and exceptions the step actually sees. That is a judgement the gate holder makes against the record rather than a calendar setting.

AI Opportunity Audit

Decide what the first gate should open on

Pick the workflow you would want to watch before you trust it, and find out whether it leaves the record shadow mode needs.

Book an audit