Five agents per engineer makes review the only human contact with Daytona's code

Each agent gets its own computer. Daytona's documentation describes sandboxes with dedicated kernels, filesystems, network stacks and allocated compute ^[12]. The agent creates that environment through an API or an SDK, runs processes, preserves state and resumes work from a snapshot ^[13]. The isolation exists so agent code can execute and use development tools without exposing the customer's underlying infrastructure ^[14]. An operator running five of those is not typing. The job is specifying the work, then reading what comes back and correcting it.

Sixteen engineers at five agents each is about 80 sessions in flight ^[1], inside a company of roughly 30 people, so engineers are a little over half the headcount ^[2][2]. Generation scales with the agent count; reading does not. Five agents can produce five times the drafts, branches or pull requests without improving what reaches production ^[15]. Burazin's figures count agent use. Daytona did not publish code shipped, defects introduced, review time or customer outcomes ^[16].

The hiring claim is about the specifying half. "People who manage people and understand how to communicate what they want" are better at running agents, Burazin told Business Insider ^[17]. He said less experienced managers can become frustrated when agents fail to produce an outcome that was never fully specified ^[18]. He also said managers without practical knowledge in their function are not useful; under Daytona's model the operator still has to understand the code or product well enough to identify a weak result and redirect the process ^[5][6]. This is a founder's observation from inside one workplace, and not a controlled comparison of managers and individual contributors ^[7].

The nearest published evidence is about instruction quality. A Columbia Business School research summary dated May 8th described experiments involving about 3,750 participants and roughly 37,000 prompts ^[8], which works out to about ten prompts per participant ^[3]. On a task with a precise target, adapting prompts accounted for nearly half of the measured improvement associated with a newer image model; on open-ended creative work the effect was much smaller ^[9]. For Daytona's pattern to transfer to another team, the work queue has to resemble the precise-target condition: an acceptance criterion someone can write down before the agent starts. The experiments did not test whether people managers outperform individual contributors ^[10].

In June, Daytona moved its production codebase to closed source, arguing that publicly exposing its isolation layer gave AI-assisted attackers a detailed map of the security boundary they would want to escape ^[11]. That isolation layer is the product ^[14]. If nobody at Daytona writes code directly ^[4], then humans specify that boundary, agents write it, and review is where a mistake is either caught or missed. Burazin and Vedran Jukic have been working on development environments since they started Codeanywhere in 2011, and Daytona began in 2023 with a different primary user ^[19].