Blog
Engineering notes from an AI-native studio — agentic web, GEO, and building with AI.
- · Alexis Grigoryev, Ivan Oparin
Instruction Markers as Diagnostics
A bounded experiment with tagged evaluation copies shows how instruction IDs can support observable diagnostics without entering production prompts or becoming claims of causal provenance.
- · Ivan Oparin, Alexis Grigoryev
Why Codex burns a weekly limit in a day while the agent waits for the tide
Spin-waiting in Codex CLI: goal mode restarts the model 0.03 seconds after every turn and demands proof of waiting by live polling, the pause tool is issued to one model in the catalog, and every poll rereads a context of hundreds of thousands of tokens. In an archive of 3808 sessions, the 52 with goal mode ate half of all input tokens. How Claude Code solves the same task, and how to describe it in async/await terms.
- · Alexis Grigoryev, Ivan Oparin
How Far to Optimize an Agent's Context
One row set measured rung by rung: minification, field selection, compact representation, the same two choices over MCP, and the abbreviated header that stops paying. With the adapter code that decides whether the backend does less work or only the agent sees less.
- · Alexis Grigoryev, Ivan Oparin
A Semantic Core Instead of a Phrasebook
What a rejected Caveman prompt rewrite taught us about keeping stable test taxonomy outside runtime context, annotating cases at evaluation time, and validating evidence before reporting it.
- · ✳ Claude Fable 5, ❋ GPT-5.6-Sol, Timur Kackan, Ivan Oparin
From AI Concept to Production Identity: When AI Should Write SVG
How Relux Works turned an AI-generated direction into a complete, reproducible identity system, with lessons on AI handoffs, SVG construction, tracing, Inkscape, and production QA.
- · Ivan Oparin, Alexis Grigoryev, Timur Kackan
How We Made Our Site Agent-Native, Without Cutting Corners
The story of relux.works going from an agent-readiness score of 29 to a perfect 100, Level 5: a working MCP server, an inquiries API, OAuth for agents, five registries, the checks we deliberately left red, and a sober look at what a perfect score is actually worth.