MENGENAI EPISOD INI
Kiriti was an early member of the Codex team at OpenAI and is now CTO at Level Up Labs. We go deep on the engineering behind coding agents: what a "harness" actually is and why it matters as much as the model, how compaction and prompt caching work under the hood (including the bug that paged the Codex team every night), how you evaluate an agent when the model keeps changing, and how he personally ships large features with parallel agents and automated code review.
If you're into AI engineering, harness engineering, prompt engineering, or loop engineering, this one's for you.
Chapters
0:00 Intro
0:32 Meet Kiriti
1:26 Prompt vs. harness
4:05 Inside a harness
6:21 Who needs harnesses
7:40 Tool search
8:40 Better harness wins
11:49 How compaction works
14:28 Server-side compaction
15:49 Evaluating harnesses
18:21 Prompt caching
21:56 Front-end taste
24:58 Monitoring live sessions
27:05 Learning from traces
30:23 Dropping in GPT-7
32:27 Leaving OpenAI
35:01 Startup vs. lab
36:32 Right-sizing models
38:42 Handwritten code?
39:47 His feature workflow
44:10 His agent setup
45:50 Code review now
48:55 Reviewing at speed
50:08 Getting good
53:12 Find Kiriti
Links
Kiriti on LinkedIn: https://www.linkedin.com/in/sai-kiriti-badam/
Kiriti's Substack: https://substack.com/@kiritib
Kiriti's Maven course: https://maven.com/aishwarya-kiriti/genai-system-design
Free lessons on Maven: https://maven.com/search?query=kiriti+badam&tab=free-lessons
Level Up Labs: https://levelup-labs.ai/
TRANSKRIP 🔗
speaker-1: Let's just say GPT seven comes out tomorrow. Can you just drop that right into the new codex and expect things to go well?
speaker-0: The idea of currently the code review is essentially making sure that you are sending a PR for code review only after Codex agrees with you. I mean until like GP…