+ The Blog

Notes from production AI.

Field reports from building agentic workflows, permission-aware RAG, and the LLMOps to run them. What breaks in production. What we ship anyway. What we learned.

All Engineering Operations Opinions Product

+Opinions and Expertise

The best case for agentic coding is the software you build only for yourself.

The standard objection to agent-written code is security, and it is correct for anything networked and multi-tenant. It mostly evaporates for a local tool with no server, no account and no cloud. That happens to be exactly the software you cannot buy a good version of.

by Alexander Myasoedov · Aug 22, 2026 · 13 min read

Spec-driven development is not a silver bullet. It is the new Scrum.

Writing the requirements down before the agent starts is a real upgrade over vibe coding. It is also a process, not a solution. The spec is prose, …

by Alexander Myasoedov · Aug 21, 2026

+Opinions and Expertise

Agents can finally squash decades of technical debt, if you stop treating the old code as the source of truth.

Agentic coding is at its best inside a legacy system: rewrite a service into a better technology fit, swap the database engine, carve up the monolith …

by Alexander Myasoedov · Aug 10, 2026

+Opinions and Expertise

Stop asking candidates to write code. Ask them to review it.

Models write the diff in seconds and a take-home proves nothing. The skill that separates engineers now is spotting the bug that passes CI and breaks …

by Alexander Myasoedov · Aug 5, 2026

+Opinions and Expertise

The strongest type system just lost to the fastest feedback loop.

Haskell should have been the perfect language for AI code generation: the compiler catches at compile time what other stacks need whole test suites to …

by Alexander Myasoedov · Aug 2, 2026

+Opinions and Expertise

Writing code is cheap now. Verifying it is the bottleneck.

A fleet of agents can rewrite a million-line runtime in eleven days. No team on earth can review a million lines in eleven days. The scarce resource …

by Alexander Myasoedov · Aug 2, 2026

+Engineering

Verify the delta, not the state.

An agent's eval can go green without the agent doing anything - the environment already contained the answer. Delta verification captures a baseline …

by Alexander Myasoedov · Jul 14, 2026

+Engineering

A dependency firewall for pip and poetry.

pip install runs arbitrary code as you, before any human reads the package. depwal screens every dependency on metadata alone - author history, …

by Alexander Myasoedov · Jul 11, 2026

+Opinions and Expertise

The abstraction tax: MCP versus a plain CLI.

You connect ten MCP servers to an agent and it has spent tens of thousands of tokens before you type a word. A CLI does the same job in eighty. Here …

by Alexander Myasoedov · Jul 7, 2026

+Engineering

Elwen: let an agent go find what broke.

Agents let us ship a feature in an afternoon. The testing didn't keep up. Elwen is a thin harness over your CI/CD that points an agent at your live …

by Alexander Myasoedov · Jul 7, 2026

+Engineering

How Claude keeps from calling tools in a cycle.

A model that never stops calling the same tool is the most common way an agent turns a task into a runaway bill. Here is the small machinery that …

by Alexander Myasoedov · Jul 6, 2026

+Engineering

The blackboard pattern for multi-agent memory.

A design pattern from 1970s speech recognition turns out to be the cleanest way for several agents to collaborate without knowing about each other.

by Alexander Myasoedov · Jul 5, 2026

+Field Reports

State of AI dev - June 2026: the month the demo stopped counting.

A field report from the work we did this month. The industry quietly agreed on one thing in June: a prototype that works is no longer evidence of …

by Alexander Myasoedov · Jun 16, 2026

+Engineering

The agent loop: ReAct from scratch.

An agent is not a model. It is a while-loop wrapped around one - and the loop is where all the interesting engineering lives.

by Alexander Myasoedov · Jun 10, 2026

+Opinions and Expertise

The rise of skill-matched, momentum-driven code reviews.

Static PR queues were built for a world where humans wrote every line. Now the model writes the diff and the queue is the bottleneck. Here's what …

by Alexander Myasoedov · May 8, 2026