The development philosophy positions the human as a high-level architect or manager. Humans make critical architectural, product, and UI decisions, while AI agents handle the implementation, akin to a CEO managing a team or a conductor leading an orchestra.
The speaker posits that code is becoming 'sawdust'—a disposable byproduct of the software creation process. The true value is shifting to the descriptive prompts and high-level instructions, suggesting that with better models, entire codebases could be regenerated from the same prompts.
The Conductor team eschews traditional analytics and A/B testing, instead relying on their own daily use of the product ('dogfooding') to guide feature development. This 'gut-feel' approach leads to a highly opinionated tool that enforces specific workflows, such as requiring all changes to go through a PR.
The limitations of running agents on a local machine (e.g., they stop when the laptop is closed) are a key driver for Conductor's investment in cloud infrastructure. The team anticipates a near future where agents are significantly more powerful and require persistent, scalable cloud environments to perform complex, long-running tasks.
To combat the risk of AI models entering a negative feedback loop of writing poor code based on existing poor code, the team establishes 'slot-free zones'. These are human-written and -vetted parts of the codebase that serve as a high-quality foundation, ensuring the AI has good examples to learn from.
Keep pulling the thread on Charlie Holtz.