Greg Isenberg and Nick Vasilescu test the idea of treating an AI agent as a co-founder rather than a single-purpose automation. Their setup gives the agent a persistent computer, memory and controlled access to communication, research and business tools so it can carry work across multiple steps.
The demonstration moves from researching startup ideas to building a landing page, shaping an offer, preparing a thumbnail and drafting an outreach sequence. Greg Isenberg and Nick Vasilescu argue that an agent becomes more useful when the surrounding harness supplies relevant context and connectors, although every additional permission also increases the need for deliberate access controls and observability.
Greg Isenberg and Nick Vasilescu also run an informal side-by-side task with Grok 4.5 and GPT-5.6 Sol. Both models produce a usable page, while Grok finishes first in this single test. They present the result as evidence that speed, token efficiency and tool execution can change the practical experience of an agent, not as a controlled benchmark of overall model quality.
Watch on YouTube
