# CoreSpeed Blog

Source: https://corespeed.io/blog

Notes from building the harness AI agents run on — connectors, memory, budgets, billing, and the engineering in between.

## Posts

- [How to turn CoreSpeed into a content factory](https://corespeed.io/blog/how-to-turn-corespeed-into-a-content-factory) — 2026-09-17, by Spencer Zhao. How we made PRODUCTION INCIDENT, from Blender rehearsal and generated performances to screen compositing, sound, and the final 4K edit. (Markdown: https://corespeed.io/blog/how-to-turn-corespeed-into-a-content-factory.md)
- [A letter from our founders](https://corespeed.io/blog/a-letter-from-our-founders) — 2026-09-14, by Milton Yan, Hao Su. Why we're building an operating system for agents, and what becomes possible when they have one. (Markdown: https://corespeed.io/blog/a-letter-from-our-founders.md)
- [Introducing CoreSpeed](https://corespeed.io/blog/introducing-corespeed) — 2026-09-14, by Hao Su, Milton Yan, Spencer Zhao, Benji Huang. An operating system for agents. (Markdown: https://corespeed.io/blog/introducing-corespeed.md)
- [How we took LongMemEval from 80.0 to 94.2 without touching retrieval](https://corespeed.io/blog/how-we-took-longmemeval-from-80-to-94-without-touching-retrieval) — 2026-08-27, by Spencer Zhao. LongMemEval-S climbed from 80.0 to 94.2 while retrieval recall at 10 stayed at 0.99. The gains came from evidence budget, reader generation, and instructions. (Markdown: https://corespeed.io/blog/how-we-took-longmemeval-from-80-to-94-without-touching-retrieval.md)
- [We retrieved the memory, then dropped the answer](https://corespeed.io/blog/we-retrieved-the-memory-then-dropped-the-answer) — 2026-08-22, by Spencer Zhao. A 200-question conflict run scored 0.800 at the parent-Memory boundary and 0.635 at the evidence boundary. The gap was in what reached the reader. (Markdown: https://corespeed.io/blog/we-retrieved-the-memory-then-dropped-the-answer.md)

RSS: https://corespeed.io/blog/rss.xml