Quiet, technical — agents that hold up.

I'm Nick Krzemienski. I live in New York and I've spent most of my career on the unglamorous middle of the stack — iOS apps, video pipelines, streaming infrastructure, the kind of systems that fall over in interesting ways at 3am. About two years ago I started working with AI agents the way most engineers do: one prompt at a time, mostly skeptical, occasionally impressed. That's not where I ended up.

What I watched happen — and what this site documents — is the slow shift from a chat box to an actual development practice. I went from typing prompts into a window to running fleets of agents in parallel worktrees, with consensus voting deciding whether code ships, evidence files gating completion, and skill libraries that carry context across thousands of sessions. The numbers behind that shift are specific: 23,479 Claude Code sessions across 27 projects, mined and analyzed; four production tools that came out of those sessions — Anneal (planning that hardens itself through tournament rounds), ValidationForge (no-mock evidence-gated validation), Multi-Agent Consensus (3-agent unanimous voting before merge), and Crucible (the runtime that enforces all of it); and a 32-post series writing down what actually held up under load.

The thing I keep coming back to is how much of agentic development is just discipline rendered into infrastructure. Evidence files instead of vibes. Validators instead of "looks good to me." Three agents that have to agree before code lands instead of one model confidently shipping a regression. None of it is magic. All of it is reproducible, which is why every claim on this site links to a real session, a real repo, or a real screenshot.

I'm currently available for consulting with teams shipping production AI products — particularly groups wrestling with how to move from "the demo works" to "we trust this in front of customers" without the whole thing collapsing into a tangle of brittle prompts. If that's where you are, the fastest way to talk is the consulting form at /work. Otherwise you can reach me on LinkedIn, see the code on GitHub, or find me as @krzemienski on X.