Guardrails, not manual review
You set the tests and limits once. Then agents ship without you checking each result.
Konstantin's verdict: Trial
I no longer read every result my agents produce. I set the guardrails once, and the agents ship inside them.
Automatic tests run before anything goes live. Before each release, tests click through a test copy of the product. For each bug I report, an agent first writes a test that fails because of the bug. Then it fixes the code until the test passes. A Haiku agent clicks through each change as a user would.
Software checks risky agent actions before they run, such as sending a message or reading a private file. Agents cannot turn that off. Before each release, a reviewer agent checks the change. It is Sonnet 5.5, a different Claude model that starts fresh.
Five reviewer agents read each edition of this radar as a visitor would. It ships only when all of them pass it. It sits in Trial because I stopped my own review this week. A mistake that no check covers can still ship.
What moved
On other radars
No community radar places this yet.
Sources
No outside source. The verdict comes from Konstantin's own work.
Share and embed
Makers: show where your product sits. The badge updates when the radar does.
Share this entry
Also in Practices · Trial
- Check every path through a user flowDrawing a user flow, such as a booking, as a chart of steps. A program then tries every order of steps, including retries and delays, to find where the user flow breaks.
- Give agents narrow keysGiving each agent the narrowest access key and spending limit that does the job. One mistake or breach then reaches little.
- Let agents reach you by pushAgents bring questions and results to you, by text message, phone call or podcast, instead of leaving them in notes you must open.
- Pick a direction before agents build creative workFor creative work, agents first show several storyboards or directions. You pick one, then they build a rough version for fast feedback before any polish.
- Scheduled agent runsAgents run on a schedule and work through your inbox and task board. They fix what they can. You get only the decisions.