Software architecture · LinkedIn adaptation
What made three coding agents work in parallel
A small architecture-memory project put Git worktrees and agreed interfaces to the test.
Ravi tried Herdr while making a tool that reads architecture documents and remembers decisions. The project had a memory layer, retrieval and API code, and a web interface. He gave each of those areas to a different coding agent.
Separate working copies and agree ownership
Git worktrees gave each agent a separate checkout. Ravi also specified which folders each agent could change. Those folder boundaries were instructions to the agents. The team still needed to check them at review time. Herdr made it easier to see which agent was working, idle, or waiting for input.
Agree on the interface before starting
The agents shared a data model, API specification, and sample data. That let the web interface use fixtures while the API was still being built. Ravi assigned the slower local model to the interface because the other work could proceed without waiting for it.
Start with one boundary
Before adding more agents, pick two pieces of work and write one representative request and response at their boundary. If both can proceed against that example, try the split and review the integration early.