Stories from building CRUX with its own workflow, and how teams can work with AI agents without losing engineering discipline.
Oct 28, 2026
"Assign a task, get a pull request" sounds like the end of meetings and specs. In practice the agent builds whatever the task says, guesses included, and you find out at review. CRUX takes the other road: agents that know the project, ask the person who knows, and check in with people at the points where a wrong turn is still cheap to fix.
Oct 23, 2026
Every new agent session makes you decide what to write, which skills to load and what to check afterwards. A workflow makes those decisions once, per task status, so you can launch sessions in parallel from ready-made templates, one click each.
Oct 23, 2026
Our wiki is an Obsidian vault in git plus a few agent skills. It is the reason a fresh Claude Code session knows how our system works before it reads a single line of code.
Oct 23, 2026
A step-by-step guide: an Obsidian vault in your git repo, a rules file for the agent, a map page, and a habit of keeping it true. Copy the files, run two prompts, and your agent starts every session knowing the project.
Oct 23, 2026
RAG picks text for the agent by a similarity score. For agency teams building sites and apps, we give the agent a wiki, a tracker and a code graph and let it choose what to read. Why, and when RAG still wins.
Oct 23, 2026
PMs, QA and even your clients work in the text of tasks, not in code. Advisor agents meet them there: they read the project's wiki and code but never change them, ask sharp questions, and post the answers to the task only after a person says "ok".
Oct 23, 2026
The person with the answers is usually the client, and every question to them goes through calls, emails and meetings. With CRUX the client works in the same tasks as your team: an AI agent asks them one plain question at a time and records the answers on the task for everyone to build from.