Harness bets the next platform-engineering purchase is governance for coding agents

Harness CEO Jyoti Bansal told the company's {unscripted} NYC 2026 audience that application developers may soon be creating 15 to 20 pull requests a week ^[4][9]. Over a five-day week that works out to three to four a day per developer ^[15]. Ten developers on that pace open 150 to 200 a week ^[16]. That queue, not the coding agent, is what Harness says Software Factory is for ^[1].

Harness says a specification fixes how an application gets developed so agents do not reinvent the workflow every time, and applying controls that way also reduces tokens that would otherwise be consumed ^[5]. devops.com describes the token effect as a potential lowering ^[2]. A spec that stops an agent re-deriving your pipeline on every run should cut input tokens. Harness put no measured figure on it ^[2]. Whether it cuts them enough to show up on an invoice is the part a buyer has to test after signing.

The numbers Bansal did give describe the index. The Harness knowledge graph tracks 567 entity types and takes 29 billion context updates a month ^[8]. Divide by the seconds in a 30-day month and that is roughly 11,200 updates a second ^[14]. It measures how much state the graph ingests. A platform team needs the rework rate on agent-authored changes and the time from a pull request opening to merge before it can price any of this.

Vibe Mode sits outside engineering's reporting line. It applies Harness development policies to code generated by citizen developers using AI tools ^[3]. Platform engineering ends up holding the exception queue for people who do not report to it.

Bansal's case for a control layer outside the coding tool is that AI coding tools lack the context required to successfully deploy code in a production environment ^[10]. That claim describes today. The scaling claim is a forecast, and devops.com reports it is not clear at what pace DevOps teams are injecting AI agents into their workflows in the first place ^[11]. The stated goal is to remove the scripts engineers currently write and maintain to construct a workflow ^[13].

Flex Pricing changes what a team commits to: a pool of credits spent across units of the platform instead of a licence per module ^[6]. That suits a team that cannot predict which parts of the platform it will use most. It also turns sizing into a guess about how many credits four core agents and their sub-agents will burn across delivery, security testing, runtime security and cost management ^[7].

Whether this belongs in this year's budget comes down to how many agent-authored pull requests merged last month without a human editing them, and what share of your policy failures came from someone outside the engineering org. High on the first and the spec layer is buying throughput against a baseline you already hold. Near zero on both and you are governing a workload that has not arrived yet. devops.com did not report a ship date for Software Factory or a price per credit ^[17].