The agents work in the chat apps teams already use. Buzz and CircleChat are open-source chat apps where people and agents share a channel and pass work back and forth. The same agents also run in Claude Code, in a terminal. Each agent is built from portable skills, so it works the same way on every platform.
Turns interview transcripts into themes and findings. She does it by handing each step to one of five sub-agents.
Sub-agents reporting to Claire
Scribe Splits each transcript into lines and stores them.
Lexicon Finds the key terms, phrases, and concepts in the transcripts.
Tagger Tags each transcript line with its sentiment.
Analyst Writes up findings and pain points from the tagged lines.
Percy Builds personas from patterns in the interviews.
Turns a product request, like a PRD, an interview, or research findings, into a working screen built on the design system. She delivers it as a pull request with a link to a working preview. When a request is unclear or two people disagree, she asks instead of guessing, and she never merges her own work.
Checks each screen against the skills in the knowledge repo and lists every problem she finds, with the file and line number. She reports problems but doesn't fix them, and she never changes a rule.
Sub-agents reporting to Barb
Checker Checks a screen against one skill at a time, going through that skill's checklist item by item.
Feisty Tries to prove each finding wrong before Barb reports it, so only real problems get through.
Kev Does a quick first check of a screen using a model that runs locally. It flags likely problems fast, and Barb's full review confirms them.
Keeps the design knowledge accurate. He reads designers' reports from Snippy and Barb's review findings, looking for gaps or mistakes in the skills. When he finds one, he suggests a fix in a pull request, with the feedback that prompted it. If the problem is in an adapter or a theme instead, he opens a GitHub issue on that repository. A person reviews and merges every change, and Alan doesn't build screens himself.
Writes realistic fake interview transcripts, so we can test and demo the research pipeline without using any real person's words.
Runs the review app where a person checks each finding against the exact quote it came from. A finding only goes live after a person approves it.
Agents are only as good as the system they build against. If your tokens live in three places and your rules live in people's heads, an agent will guess, and it will guess differently every time.
Take the quiz