Plasmidsaurus CTO Swizec Teller reports that AI agents now write 97% of the company’s code, fundamentally altering the engineering workflow at the 21-person biotech startup. By leveraging Cursor’s background agents and strict accountability policies, the team ships hundreds of pull requests weekly, with 16% automatically approved and merged without human intervention. This velocity, once unimaginable in operations-heavy environments, has shifted the primary engineering challenge from code generation to rigorous quality control.

The Shift to Delegated Engineering

The transition began in July 2025 when Cursor launched background agents, allowing engineers to delegate entire tasks rather than supervising AI in real-time. Teller describes this as moving from 'vibe coding,' where developers watched the AI think, to true delegation, where agents produce code, screenshots, and demo videos for asynchronous review. This approach frees senior engineers from the friction of prompt engineering, enabling them to focus on high-level architectural decisions and user impact rather than syntax.

Accountability and the Phone Number Policy

Despite the heavy reliance on AI, Plasmidsaurus enforces a strict 'we have your phone number' policy, ensuring every engineer remains personally accountable for the code they ship. There are no translators between business needs and code; engineers must own features from concept to production, engaging directly with stakeholders. This culture of high agency prevents the diffusion of responsibility that often plagues AI-assisted development, ensuring that speed does not come at the cost of operational stability.

Code Quality as a Financial Metric

Teller highlights a critical economic insight: poor code structure directly increases token costs for AI agents. In one experiment, improved file organization reduced token consumption by 83% for a specific coding task, proving that architectural cleanliness is now a measurable financial metric. Because agents amplify existing patterns, a messy codebase leads to expensive, inefficient iterations, while well-structured modules allow agents to navigate and modify code with greater precision and lower cost.

Review as the New Bottleneck

With code generation becoming nearly free, the bottleneck has moved to review, where human judgment remains irreplaceable. Teller notes that automated code review tools often miss context-specific nuances, leading to a high rate of irrelevant suggestions. Senior engineers now spend significant time reading and refining AI-generated code, turning manual feedback into deterministic linters and documentation rules. This process creates a feedback loop where agents learn to adhere to internal standards, gradually reducing the need for human intervention in routine changes.

Key Takeaways

  • AI agents now write 97% of code at Plasmidsaurus, with 16% of PRs auto-merged.
  • Background agents in Cursor allow engineers to delegate tasks rather than supervise generation.
  • Poor code structure increases token costs by up to 83%, making architecture a financial priority.
  • Human review remains the critical bottleneck for judgment and context, not code creation.

The Bottom Line

The era of writing every line by hand is ending, but the need for senior judgment is intensifying. If you aren’t building systems that enforce quality and context, you’re just burning tokens on slop.