How we work with AI agents
Tishi is a team of five that builds Moss together with a number of AI agents. This post describes how the work is divided and the rules we follow.
Tishi currently has five members. Most of Moss's code, tests and release process, and this website, were built by the team working with a number of AI agents. This post describes how.
Division of work
- People set direction. What to build, for whom and to what standard is decided by members of the team.
- Agents carry out the work. Writing code, running tests, research, comparing options and organizing documents are done by AI agents, often in parallel.
- People accept the work. Everything we release has been used and confirmed by a member of the team.
Rules
We have set the following rules for agents:
- Work that has not been verified must not be marked as done.
- Secrets must not appear in output, commits or logs.
- Each task uses its own branch and is reviewed by another party. The agent that writes a change cannot merge it.
- Anything involving spending, publishing or production requires explicit consent from a team member.
- Tests must not use real users' data.
How this relates to the product
This way of working is also what Moss aims to offer its users: one person directing several agents to complete work. We use it in our own work first, and apply what proves effective to the product.