About Mihir Wagle.
By day, I am a technologist at Microsoft; by weekend, I am a builder, hiker, and enthusiast of detailed universes.
I am a proud graduate of IIT Bombay, a cricket fanatic (ask me about any T20 match and I will have an opinion), and an overthinker by nature.
This blog is a collection of my notes on the things that occupy my mind — ranging from the future of agentic AI and T20 cricket to the best historical sites to visit on a family trip. I value specific metrics over vague claims and clear mechanisms over corporate-speak, a philosophy I try to bring to both my professional work and my personal writing.
Things I build on weekends
Mostly at the intersection of agentic AI and cricket — the two things this blog can't stop talking about.
Daftari (دفتری) is the Urdu word for a ledger-keeper — corrections were noted, never erased.
A long-term memory cortex for LLM agents: a plain-markdown vault they read, write, and curate over time — portable across any model, on your disk, under git. Contradictions surface as tensions, held open rather than flattened into a false answer: a tension may never masquerade as a supersession.
A cricket stats nerd’s dream, wired directly into Claude.
An MCP server that turns 10.9 million ball-by-ball deliveries from Cricsheet into a queryable cricket brain — 21,000+ matches in a local DuckDB, answering questions like “How does Kohli bat against Hazlewood in ODIs?” in milliseconds.
Agents do not need more trust. They need a boundary.
Runs tool-calling agents inside an explicit envelope — what they may read, write, and spend — with every run reviewable by a Third Umpire. For when an agent is useful enough to delegate to, but not safe enough to leave unsupervised.
A series of two-way doors composes into a one-way door.
A research prototype for path-level agent safety: a trust budget, staging area, and irreversibility classifier that catch harm composed across many individually-reasonable actions — the failures no single forward pass can see.