Chapter 7 of 11 · ~2 min

A person, when it matters

Below fifteen percent the quote goes out on its own. Above it, a person decides. That sentence is the whole design of the human step, and the workflow makes it literal: a Conditional on the price step's needs-approval field, and on the true branch an Input for a second participant, the rep.

The rep's step. The rep receives it with everything they need: the company, the quantity, the requested and permitted discount, and what the research found. It is an assisted input, so the rep can ask a model to summarise the research or explain the pricing before deciding, and the model is told never to decide. It has a twenty-four-hour timeout that escalates to the sales lead. The answer is one of two buttons. The decision is recorded under the rep's name, next to what the model proposed.

Two participants, and what that changes. The prospect is on the widget on the company's website. The rep is a member of the sales team. Because the workflow has two human participants, it cannot be a public deployment: the platform allows anonymous starts only for single-participant workflows. It is deployed privately, the website's backend starts each run and mints the prospect's token, and the rep is bound to the team at run time. Chapter 9 shows the code.

The recorded run. Harbour Hotels asked for twenty percent on sixty mid-range kettles. The checks passed, the research came back with fit unknown because the search failed, the price step granted fifteen and flagged the request, and the run paused for the rep. Twenty-one seconds later the rep approved. The decision step then recomputed the total at twenty percent, the quote went to the prospect, the deal was written to the CRM, and the run entered a seven-day wait. Twelve steps, two model calls, and one decision by a person, all in the record.

Experiment

Runs in your browser

The recorded run, replayed in the inspector. Click through the steps; the rep's approval is the one with a person's name on it.

Run 64dbc007, a test run of the private staging deployment, driven by the coding agent as both participants. Twelve steps from the form to the seven-day wait; one minute of wall-clock time.

Agent · agent · research

research

Input
Instructions with the four ranges; the enquiry; existing customer: false. One tool: web_search. Output schema with five fields.
Execution
Call 1: 931 tokens, 3.4 s, requested web_search. The search tool returned an error. Call 2: 4,374 tokens, 18.3 s, final answer. Schema check passed first time.gemini/gemini-2.5-flash · 5000 in / 305 out · 21.8 s
Output
fit: "unknown" · reasons: "…could not be retrieved due to a failed web search…" · evidence: "Web search for Harbour Hotels in the United Kingdom failed. Based solely on the provided information: …" · suggested_products: "Birch" · questions_to_ask: one question about room aesthetics
State write
research = "{fit: \"unknown\", reasons: …, suggested_products: \"Birch\", questions_to_ask: …, evidence: …}"

Two model calls in twelve steps, both bounded, neither touching a number. One decision by a person, recorded under their role next to what the model proposed. Every other step is a rule, a fixed call or a message, and the whole run can be replayed exactly as it happened.

Big question

Where is the line in your process above which a person must decide, and is it written down as a number?