Mihir Kachroo

wrapped up internship #5 at databricks last week. probably my favourite one so far. got lucky to work with a great team on cool problems alongside some really awesome people

Can agents learn from past successes — without fine-tuning? In our latest blog at Tile Labs, we test a simple idea: giving agents examples of successful past runs before they act. No weight updates. No retraining. Just better context. What we found: • More selective agents → fewer errors • Smaller models improve the most • Frontier models gain better calibration Read more: tilelabs.dev/blog/implement…

🚀 Our first blog is live: Benchmarking AI Agents using RL Environments We evaluate frontier models and uncover a key insight — the gap between completing a workflow and completing it correctly is larger than you think. Read more: tilelabs.dev/blog/benchmark…

We’re excited to introduce Tile Labs. A research lab focused on advancing model intelligence — exploring how AI systems become more capable, adaptive, and autonomous. More to come soon. hello@tilelabs.dev

better models matter but better harnesses — the infrastructure around them — may matter just as much openai.com/index/harness-…