How Ryan Carson Runs a Startup with AI Agents

Founder Ryan Carson joins Peter Yang to describe how he uses AI agents while building his startup. He emphasizes that setup is still difficult and that his approach depends on written context, repeatable workflows, and ongoing maintenance. The episode contains a WhisperFlow sponsor segment.

Ryan’s OpenClaw assistant, R2, uses a set of instructions and a priority map to help with scheduling, inbox checks, tasks, and business research. He describes recurring checks and communication through Slack. His own experience with scheduling reliability is specific to the setup and product version he tested.

The assistant runs on a separate machine, which Ryan accesses remotely. He uses a coding agent to inspect configuration and troubleshoot problems when the assistant fails. A second tool can help diagnose issues, though changes still require appropriate review and a way to recover if they fail.

Ryan shows how a scheduled task invokes an executive-assistant skill and reads longer-term priorities. He gives the assistant a distinct email and GitHub identity and selected access to his own calendar and inbox. The exact permissions determine what it can do; a separate account alone does not guarantee safe behavior.

He says he is delaying hiring while learning business functions himself with help from agents, despite having raised funding. His comparisons between agent and human onboarding are personal judgments, not measured evidence that one can replace the other across roles.

Ryan describes a scheduled prospecting workflow that researches potential contacts and tracks outreach in a spreadsheet. In his setup it can send messages; he discusses instructions to avoid duplicate contact and an email behavior he had to investigate. External communications and access to sensitive data need clear authorization, careful controls, and monitoring.

For weekly priorities, Ryan recently returned to a paper list because it gave him a better view of the week. He also tries to balance customizing his assistant against the time spent maintaining tools, preferring an existing platform with selected skills and scheduled tasks.

For development, Ryan uses cloud environments and playbooks. He describes scheduled browser tests that exercise an end-to-end user journey and record results. Automated testing can provide useful evidence, but a successful run does not establish that every failure mode or security issue has been covered.

His feature process begins with a spoken description and an agent-assisted requirements document. He then delegates implementation and testing, reviews feedback, and uses a playbook to land a pull request. His reported shipping rate and preference for automated review are individual experiences; teams should choose review and release controls for their own risk.

Ryan also describes using agents for advertising analysis and content production. He says he first paid a designer for brand work and reference images, then used those assets to guide subsequent content. He plans eventually to hire people for functions he has learned to operate himself. The speakers conclude that setting up and refining systems can be useful, while recognizing that it takes substantial work.