The two things a founder still has to do when agents run the install

On 2026-09-10 we rehearsed the whole path a client takes through our operating system, on our own stack, with the model mocked and the vendors in fixture mode. A text arrived. 808 seconds later, thirteen and a half minutes, a workflow was live and measuring its first missed call. The log records every step. Inside the walk, two of them were a person at the firm. The rest were the client, software, or in one case the test harness nudging a link the mock mailbox had swallowed.

The two it did not do are the subject of this piece. They are not gaps still to close; they are the two places where the system is built to refuse.

The first: the discovery call

When a lead texts, the Sales agent is built to read it, check it against the fit rules the owner set, draft a reply, and book a discovery call. Every one of those steps is software. The call itself is not.

The deal cannot move past the call until a person marks it delivered, and the state machine refuses an agent as the actor on that stage. In the rehearsal log the step reads "run the discovery call and mark it delivered", and the actor column reads "human".

The reason is not sentiment. The call is where the scope comes from. What is costing this business the most time and which install answers it are judgments about a person's business, made with the person, and we do not think a model should make them alone. After the call one agent adds the install to the deal and another drafts the agreement from those lines.

The second: the countersign

The client receives the statement of work by email with a one-time link and accepts it on a page of ours. Their name, email, network address and the time go on the record. From that moment the agreement's text and terms are frozen; the database refuses any change to them.

Then the owner countersigns, and only the owner can. The rule lives in the database, in the trigger that guards the contracts table. Its refusal reads: only an owner in a signed-in session may countersign; agents and services never sign. Neither a service account nor an agent acting under a signed standing order can do it. Signing as the owner is also on the list of actions no tool at the firm may ever be given, a list that is code and that shrinks only on the owner's written instruction.

The countersign is what opens the engagement, raises the install invoice and starts the install clock. The one action that commits the firm to a client is the one action only a person can take.

Everything else, with a stop button

Almost everything between those two touches is done by agents, and "done by agents" does not mean unsupervised.

An action an agent takes that would reach the outside world or move money, a text, an email, a charge, goes into a queue first. It waits thirty minutes by default. During that window a person at the firm can stop it. In the rehearsal the window was one minute so the walk would finish in an afternoon; the log says so on each of the eight queued actions, and every one fired on its own with nobody approving it.

The window is not permanent. After an agent has executed the same kind of action twenty times without anyone stopping it, that action runs at once. One stop resets the count to zero. And no setting can shorten the window to nothing: a flag set below the minimum falls back to the default.

There is a third human hand on the path, and it is the client's. No agent can take a workflow live before the client has said yes in writing, and the reply that records it must come from the person the question was sent to. The one person who can override that gate is the owner, in writing, and the override is recorded in the audit trail where the client can ask to see it. In the rehearsal the client's "YES" is step forty-one.

Why we built it this way

We think a founder who hires an implementation firm is not buying a promise that software will run their business. They are buying a person who will decide what goes in, sign for it, and stay to run it. Our system is built so that the person cannot be skipped at exactly those two points, and has a window to step in at every point that is new.

That is the deal. Software does the steps that are steps. A person does the two that are decisions.