How I work with agents
- Shipping
- Coding agents do most of the typing. My part is the harness: the rules they follow, and tools that let them read the running system instead of guessing.
- Products
- Agent loops with tools, memory and a hand-off to a human — a cheap classifier first, a chat model only when it's needed.
- Trust
- Evals on real scenarios before a prompt or model change ships, and the spend checked before every call.
Before agents: 13 years of backends — payments, ad-tech, data platforms. The whole history is on LinkedIn.
Ask your agent about me
claude mcp add --transport http kondaurov https://kondaurov.dev/mcpA read-only MCP server: profile, posts, search and live project stats. Any MCP client takes the URL https://kondaurov.dev/mcp; every page also answers Accept: text/markdown.
Things I've made
- procdeckGoRuns my whole dev stack in one browser tab and hands it to the coding agent over MCP — so it debugs from what happened, not from re-reading code.
- effortless-awsTypeScriptI got tired of CloudFormation, so I wrote a framework where the handlers are the infrastructure: write the code, run one deploy.
- tg-bot-sdkTypeScriptA Telegram bot SDK typed straight from the Bot API docs, with an in-memory emulator for tests. No runtime dependencies.
- Typed AWS SDKEffectGenerates Effect wrappers for any AWS client you install — every documented exception becomes a typed error.
Writing
- The bill for free software
- Coding isn't dead. It's where the loop closed first.
- A harness for an AI-native app: evals, labs and a fast judge
- Coding agents: the boost is real, and so is the plateau
- Shipping a dev tool as one binary
Also a job-search guide in 6 chapters · all posts