Vercel's inbound agent now runs on 14 rules, not 1,000 lines
Vercel's COO has described how one engineer at 20% time, a prompt written by the best SDR and six weeks of human checking produced an agent that handles the top of the funnel. The lesson is what came after: most of the prompt turned out to be code.
Vercel handed its inbound qualification problem to a single engineer on a new go-to-market engineering team, working on it for roughly a fifth of his time.
The first version was not written by the engineer. It was a prompt of about 125 lines written by the best sales development rep on the team, encoding the company's rules for qualifying a lead.
That is the first transferable move. The person who already does the job writes the instructions; the engineer builds the plumbing.
In phase one the agent did everything except press send. It researched accounts, qualified them and drafted the outreach, and a human read the output.
Jeanne DeWitt Grosser, Vercel's chief operating officer, has compared this to onboarding a new hire: read all of the first hundred emails, then drop to a random sample of one in a hundred.
The company started in June and removed the human in August, according to her account on Tomasz Tunguz's Office Hours podcast. Six weeks of supervised output was enough.
Then the thing got fat. Over the following year Vercel added product surface area and moved upmarket into enterprise, which brought a wider spread of inbound companies, and the prompt grew to around 1,000 lines.
Grosser's complaint about that state is the practical one. The model would not reliably follow every rule buried in the prompt, and a good deal of qualification is not a judgement call at all.
So the team split the prompt in two. Engineers wrote the deterministic parts as code and left the model only the decisions a human would previously have thought about.
Inbound now runs on 14 rules, on the company's account. When a lead arrives the system checks Salesforce for an open opportunity on that account and, if there is one, routes the lead to the account executive. No model opinion is invited.
Because the rules are written down, a second agent can watch for breaches. When one of the 14 is broken, that agent decides whether the breach was justified, and either fixes it or lets the exception stand.
Tunguz reports the same shape elsewhere: across 14 production agent workflows he examined, 65% of nodes run as pure code and only 14% remain fully agentic.
The staffing consequence is worth noting. Vercel's SDR team were all promoted to outbound, skipping the usual year in the role, on the argument that humans are better used talking to humans than writing email.
Tunguz puts the running cost of the whole function at about $1,000 a year in inference and infrastructure. That figure is his, not a published Vercel number.
For a founder, the sequence is the point. Pick a problem that is mostly deterministic, let your best operator write the first prompt, supervise it like a new hire for six weeks, and then keep auditing which lines of that prompt were never judgement in the first place.