About
I’m Amit Agrawal. I carry revenue and retention numbers for a living, and I build the infrastructure underneath them.
That started in the datacenter, designing compute clusters and private clouds. It never really stopped — the substrate moved to public cloud and the workloads became agents. Identity and authorization boundaries, managed runtimes for stateful sessions, gateways in front of tool APIs, retention and cost controls: that’s still the layer I work at, and most of what ends up here.
Most people pick one side: the systems, or the outcomes they’re supposed to produce. I’ve been accountable for both the whole time. The interesting problems live where they collide.
If you’re wiring agents into real workflows and hitting cost, auth, or memory walls, that’s most of what I write about.
This site is the build log — 80 posts since March 2026, nearly all of them from systems I run daily rather than demos I set up in order to write about them. A cross-app agent task that cost $5.44 on the naive path and $0.17 once both legs were optimized, a 32× spread that came from implementation choices rather than model choice. A 186-request test showing prompt caching returns 84.4% input savings on a stable prefix and costs 25% more than uncached when the cache key drifts. An AI code-security tool pointed at my own control plane with a planted IDOR bug, to find out whether it would actually catch it.
The method is the same every time: move from a specific problem to a working system fast, then generalize the pattern once it survives real conditions. Conviction first, then build. Proof over slides. Most of what’s here started as something I needed for myself — cost discipline in an experimentation account, graceful lifecycle for stateful workloads, making agent patterns reusable — and got written up once it earned its keep. I wrote about why this isn’t a new identity in Builder Is an Operating Mode, Not a Job Title.
If a workflow or tool isn’t AI-native, I usually won’t invest in it. The friction and the ceiling are both too high.
Where to start
- API → Connector → Browser → Computer Use → Human — a cost-justified hierarchy for agent tooling, with the numbers behind it.
- Skills Are a Git-Native Distribution Primitive — why agent capability should ship the way code ships.
- Prompt Caching Is a Harness Design Problem, Not Just an API Toggle — where the savings actually come from.
- How this site is built — infrastructure choices, syndication pipelines, and the ongoing building-in-the-open experiment.
New posts go out over RSS. That’s the one thing worth clicking.
Disclaimer
All views and content here are entirely my own. Nothing represents my employer or any affiliated entity.