Boris Cherny: Building Claude Code

Successful agent-based product development now relies on an empirical, iterative process where developers remove restrictive instructions and allow models to handle complex, long-running tasks autonomously, using test-time compute to verify results.

• Model longevity — Opus 5 functions autonomously for extended periods, from days to months, without needing manual scaffolding [01:43].

• Reduced restrictions — Newer models possess internal safety features that make them resistant to unauthorized command injection, removing the need for defensive, bloated prompts [02:35].

• Ablation process — Developers should periodically delete system instructions to identify what is actually required for the current model, preventing legacy constraints from hindering performance [06:03].

• Empirical observation — Rather than predicting failures, run the software, observe where the model struggles, and add instructions only when a recurring issue is identified [08:16].

• Product overhang — Many AI tools fail to utilize the full capability of current models because the software harness creates artificial limitations [11:05].

• High-level steering — Performance improves when users define clear goals and exit criteria instead of providing granular, step-by-step commands [15:20].

• Dynamic orchestration — Complex goals are achievable by triggering thousands of agents that manage tasks in parallel, effectively scaling test-time compute [17:34].

• Self-maintenance — Models can independently manage code bases by identifying dead code, fixing redundancies, and proposing daily updates [28:33].

• How do dynamic workflows manage parallel agent tasks within a sandbox?

• What is the role of verification when assigning high-level goals to an agent?