From the work
Notes
Practical findings from building Kogen.
Build your own harness
Why we're building a harness around Kogen's workflow, what the newer Codex comparisons show, and which changes haven't helped.
Read the noteTuning macOS for development
What I changed on my Mac after slow Kogen builds led me to Spotlight, fseventsd, XProtect and automatic Xcode updates.
Read the noteWhy I fork my ChatGPT conversations
Organized project files already save me from repeating workflows. Forking a prepared conversation lets the next task reuse the reading too; caching was what led me to investigate it.
Read the noteHow much spec does an agent need?
What our benchmarks taught us about giving coding agents a useful target and checking what they build.
Read the noteDon't get locked in
I've moved between Claude and Codex. I want the skills and setup I build to move with me when a model, subscription or harness stops fitting my work.
Read the noteJev this, Jev that
The boring Jev uses I've found in Kogen: reading an agent's notes and checking requirements before coding.
Read the noteCheck what your agent actually receives
In Optimum Codegen, 31 rules we had written for the build agents reached them as one. Here is what broke the delivery, why nothing caught it, and how to check your own setup.
Read the noteReplacing Codex without replacing Kogen
I want to try other coding agents without handing them Kogen's checks, review, history, and definition of finished.
Read the noteI don't want AGENTS.md
One repository-wide instruction file cannot fit agents with different jobs, and sharing its filename does not standardize how tools interpret it.
Read the noteMake no mistakes
Preventable mistakes enter an agent-led Build through ambiguity. Make the decisions explicit enough to inspect and check.
Read the noteMaking the offline check fast enough to keep running
A slow verification step gets deferred or skipped. Here's how Kogen kept its check fast without cutting what it verifies.
Read the noteThe agent said it ran the checks
A Stop hook makes declared checks part of finishing instead of trusting an agent to report that it ran them.
Read the noteMy Codex is not Kogen's Codex
Kogen pins a tested Codex version and runs it with its own settings and login, separate from mine.
Read the notecodex exec can exit 0 after failed checks
Codex can exhaust its attempts to fix a failing check while the surrounding codex exec process still exits successfully.
Read the noteMarkdown for agents. HTML for humans.
The source stays Markdown. Astro templates turn it into a website, so I don't hand-write HTML on every report.
Read the noteYou don’t need Astra
I tried cheaper models on the same Kogen build. They finished the job too.
Read the noteDon’t let the coding agent grade its own work
Why Kogen lets the Developer run focused tests but reserves final verification for automated machinery.
Read the noteHow to turn off Codex terminal animations
The setting for a calmer Codex terminal—and what a small startup comparison showed.
Read the note