Kratos Labs · Field notes

Production AI systems, from the work.

What we learn building AI systems that have to survive real data, real workflows, and real consequences.

§ 01 / Notebook

Latest entries

RSS ↗
  1. Stop Measuring AI Cost in Tokens

    Tokens measure one component of an AI system. The useful metric is the total lifecycle cost of producing an acceptable result at the required quality, reliability, risk, and speed.
    8 min read · AI economics
  2. Inside the Git Repository That Runs Most of Our AI Agency

    A practical look at the file-based operating system behind Kratos Labs, from lead research and planning to website operations, coding, client documents, and correction.
    7 min read · Operator OS
  3. Never Let an AI Coding Agent Loop Over Its Own Reasoning

    Long AI coding sessions can drift when one agent plans, builds, rewrites the tests, and judges its own work. This workflow separates those roles with fresh context, frozen contracts, and read-only evaluation.
    7 min read · AI coding agents
  4. Make the company workspace reliable: checks, history, time, and human approval

    Make your AI company OS reliable with narrow checks, full-history Git gates, result receipts, multi-host discovery, bounded workers, and exact-version human approval.
    13 min read · AI company OS
  5. Build the company workspace: from three files to one repeatable AI workflow

    Turn a three-file AI company OS into one repeatable workflow with canonical sources, skills, Git history, failure-tested checks, and exact-version approval.
    14 min read · AI company OS
  6. Stop re-explaining your business to AI: build an AI company OS

    Build a practical AI company OS so every model starts with the same rules, current context, output boundaries, and human approval.
    15 min read · AI company OS