April 2026 · 3 min read

Agents that draft, never send

There is a version of AI automation that quietly takes actions on your behalf and hopes you do not mind. We do not build that version.

Agents in the systems we build investigate, draft, and prepare. They get the work to the point of a decision. Then a person decides.

Make the boundary structural

It is not enough to ask an agent nicely not to act. The approval gate lives in the data model: there is simply no path for an irreversible action to happen without a human signing off. You cannot misconfigure your way around it.

  • Drafts wait for review. Edits become signal for next time.
  • Every action carries a trace, so the whole chain can be reconstructed.
  • How much to delegate is a setting the operator controls, not a default we chose.

Done right, this is not friction. It is the thing that lets a team hand real work to a system without lying awake about it.

FAQ

Common questions

The value is the work that happens before the decision. The agent investigates, gathers context, and drafts the action so a person reviews a finished proposal instead of starting from a blank page. You keep the judgment and skip the assembly.

Because a UI checkpoint is a suggestion someone can route around the moment a second entry point or background job appears. When the gate lives in the data model, there is simply no code path that produces the irreversible action without a sign-off. You cannot misconfigure your way past it.

Yes. How much to delegate is a setting the operator controls, not a default we pick. Some teams want every draft reviewed; others let low-stakes actions through and gate only the irreversible ones. The boundary moves; the guarantee on irreversible actions does not.