R13Show the product working
Pro

Agent inside the workflow

Establish the human process first, place the agent at one named step, and keep the approval and the audit trail visible.

Back to page recipes

Section sequence

6sections, in the order the visitor meets them. Bar height is the section's weight in the page.

  1. OrientHuman process
  2. DemonstrateAgent step
  3. ProveApproval state
  4. ProveContext and audit
  5. NavigateRole routes
  6. ConvertStart with one step

How the page argues

  1. 1

    Show the human process as it exists today, without the product in it.

  2. 2

    Insert the agent at exactly one step, named and bounded.

  3. 3

    Show the approval state: what the human sees before anything is committed.

  4. 4

    Show the context the agent used and the record it left behind.

  5. 5

    Give role-specific routes, because the operator and the manager buy different things.

  6. 6

    Invite adoption of one step, not of a new operating model.

What each section has to be

OrientHuman process
Today's process in three or four beats, recognizable to someone who runs it.
DemonstrateAgent step
One step where the agent acts, with its inputs, its permissions and its output made explicit.
ProveApproval state
The human review surface: what is proposed, what will change, and how to reject it.
ProveContext and audit
What the agent could see, what it did, and where that record is retained.
NavigateRole routes
Separate paths for the person doing the work and the person accountable for it.
ConvertStart with one step
Adoption scoped to the single step shown, with the rest of the process untouched.

Is this the right sequence?

Who it is for

Operators and their managers, who are weighing delegation against accountability.

Products it suits

Agentic products embedded in work that a human is still answerable for.

Not for

Fully autonomous systems where a human approval step would be theatre rather than value.

Evidence strategy

Which proof modes carry the page, in the order they land.

  1. Primary: the workflow, which shows the agent as a participant rather than a replacement.
  2. Secondary: the approval surface, which is where accountability is actually won.
  3. Tertiary: the audit record, which answers the question the manager will ask second.

CTA plan

What the page asks for, what it offers instead, and where.

Primary
Enable the agent on one step
Secondary
See the audit trail, or review the permission model
Placement
Primary action after the approval state; the governance routes sit beside it, not in the footer.

Responsive behavior

What the sequence has to do when the viewport stops cooperating.

  • The process beats become a vertical timeline with the agent step visually marked as the inserted one.
  • The approval surface keeps its accept and reject controls side by side on mobile, never stacked with reject hidden.
  • Context and audit collapse into a disclosure that is closed by default but discoverable.

Anti-patterns

The failure modes that make this sequence stop working.

  • Leading with autonomy, which triggers the control objection before any trust is built.
  • An approval step with no visible reject path.
  • Vague scope: an agent that 'handles your workflow' rather than one named step.
  • Hiding the audit story on a separate trust page.

Build spec

Pro

What an agent needs in order to build Agent inside the workflow: what each section must contain, the copy rule that keeps it honest, the slots to resolve from your product, and the order to build in.

Checking access to the build spec

Agent tags

The vocabulary an agent matches this recipe on.

  • agent-in-workflow
  • approval-state
  • audit-trail
  • permission-scope
  • role-routes