How I work
How I design teams of agents.
I design the systems AI agents like Claude and Codex work inside, and I find out why they fail. When my agents struggle, I ask them what’s going wrong, then build what they need.
These are the rules the tools come from. Each one links to where you can see it working.
01
Turn what you notice into something that runs.
Most of my ideas start with something I noticed: a tutorial that points at the screen, an agent that loses its place. The first goal is a working version tested on real input, not a perfect plan. What breaks gets written down and fixed.
02
The builder never grades its own work.
Work gets checked by agents who didn’t write it, ideally from a different model family. A second reviewer can catch assumptions the builder missed.
See it inBoss BattlesComposer
03
Investigate before you build.
Groups of agents research the problem from different angles, and another model checks what they found. The builders start from a researched brief.
See it inComposerPDF Rescue
04
Hand off facts, not transcripts.
A new agent doesn’t need the whole history. A short handoff can give it the project facts that matter next, with evidence it can check.
See it inWarmStart
05
Keep the job outside the conversation.
Long conversations get squeezed. The goal, what’s done, what’s running and what comes next belong somewhere compaction can’t reach.
See it inContinuity
06
No agent gets every power.
Searching, going online, watching a running app and changing code carry different risks, so they go to different parts with different permissions. Tools are granted one at a time.
See it inBug Eater’s designMisfit Hollow
07
Record failures as failures.
A failed run is recorded as failed, with what went wrong. A timeout isn’t called a success, and an empty trial isn’t called a win.
08
Ask “why are we assuming that?”
It’s the question that has changed the plan more than once. The original observation is kept word for word, and the smallest test that could prove it wrong comes first.
See it inLabs
See the discussion
What happens when Chad and Faye disagree?
A real written exchange between Chad (Astra) and Faye (Fable), from two different proposals to one practical compromise.
Pull up a chair →What this adds up to
Mistakes that get caught, and work that can be picked up.
None of this makes agents perfect. The goal is mistakes that are easier to catch, work that’s easier to hand over, and failures that are easier to recover from. That’s the system the tools add up to.
See the whole system →Building agent tools, developer tools or AI products? I’m looking for a role where building and explaining are the same job. I spot the problem, ship a working version, and keep the failures on the record. Message me on X or find me on LinkedIn.