Agents vs. Massive Physical Data: Redefining Control Flow
AI Engineer · youtube · 2026-07-20
When agents need to process terabytes of unstructured physical data (like dashcam video or sensor logs), the standard "loop, check, and re-derive" pattern leads to exorbitant costs and massive latency.
Core Architectural Shifts:
- Recalculation is the Enemy: When processing massive physical data, the cost of repeated reasoning is immense; traditional agent intuitions (like re-computing for verification or stuffing state into context) fail completely.
- Dataset as State: Datasets replace the context window as the unit of state management.
- Materialization and Recall First: Materializing processing results becomes the default operation, and "recall" becomes a first-class citizen. Subsequent queries should directly invoke discovered results rather than triggering new perception tasks.
Through a demo of Claude Code on raw S3 data, the speaker demonstrated how to slash subsequent query times from hours to seconds, achieving a multi-million-fold efficiency gap.
More from coding & agent
- Soft Clamp cuts tool-call overuse in multi-teacher distillation, from 13.7% to 9.0% — antgroup · 2026-07-21
- Agent harness memory loss and compaction are still a major usability problem — adityaag · 2026-07-21
- SpecJudge runs locally on Ollama to pick the right-sized AI model for your project — jokiruiz · 2026-07-21
- A developer maps out six design rules for CLIs that humans and AI agents can both use — yujiezha · 2026-07-21
- A coding-agent skill that forces ADHD-friendly, answer-first output — ayghri · 2026-07-21
- A set of agent skills for CAD, robotics, and hardware design — earthtojake · 2026-07-21