Garry Tan: Distinguishing Latent vs. Deterministic Computing Prevents Agent Failures
garrytan · x · 2026-08-27
Garry Tan shares a rule to prevent agent failures: the key is identifying where computation happens. He distinguishes between two spaces:
- Latent Space: Involves taste, judgment, and understanding vague requests, suited for models and steered via markdown.
- Deterministic Space: Involves arithmetic, SQL queries, and precise logic, requiring code execution.
Confusing the two leads to failures. For example, seating 5 people is easy for a model, but scheduling 6,000 people requires the agent to write code to track it. The model fails where we fail.
More from coding & agent
- Guiding AI to think outside the box improved translation speed by 42% — dotey · 2026-08-27
- Cursor Agent Ran 48 Hours Building a 3D Campus: More Detail, but Nearly Every Building Is Wrong — tristanbob · 2026-08-27
- Apodex 1.1 Report: Achieving Sustained, Verifiable Progress via Environment & Agentic Scaling — HeyAmit_ · 2026-08-27
- List features nearly 16 coding agents; how many does the world need? — thisiskp_ · 2026-08-27
- dhh Celebrates: GitHub Agents May Soon Stop Asking You to Drag Screenshots Manually — DanWahlin · 2026-08-27
- Clanker Cloud Allows Building Agents and Access via Curl — tekbog · 2026-08-27