Cobus, the supervisor-subagent architecture you describe became native with Opus 4.6's Agent Teams. What you built manually, a supervisor coordinating four specialized subagents, is now a built-in capability. I tested 4 agents in parallel on launch day and the pattern matched your description almost exactly. One lead agent distributing work to specialists, each maintaining deep context in their domain. The difference is that Agent Teams handles the coordination protocol automatically. Your point about the speed of building these frameworks is even more true now. Full parallel experiment: https://thoughts.jock.pl/p/opus-4-6-agent-experiment-2026
The iteration speed when Claude Code scaffolds these systems is wild. I've been building similar workflows and the number of experimental config files you accumulate is unreal. MCP configs, agent definitions, test harnesses. If you don't sort out local exclusion early you end up with a bloated .gitignore or noisy PRs. Wrote about using .git/info/exclude for this exact problem: https://blog.stackademic.com/how-to-test-mcp-servers-and-agent-configs-without-annoying-your-team-20692917488d
Cobus, the supervisor-subagent architecture you describe became native with Opus 4.6's Agent Teams. What you built manually, a supervisor coordinating four specialized subagents, is now a built-in capability. I tested 4 agents in parallel on launch day and the pattern matched your description almost exactly. One lead agent distributing work to specialists, each maintaining deep context in their domain. The difference is that Agent Teams handles the coordination protocol automatically. Your point about the speed of building these frameworks is even more true now. Full parallel experiment: https://thoughts.jock.pl/p/opus-4-6-agent-experiment-2026
Thank you Pawel for sharing the post on the parallel experiment...talking a look now.
Thanks! I hope you will find it good :)
The iteration speed when Claude Code scaffolds these systems is wild. I've been building similar workflows and the number of experimental config files you accumulate is unreal. MCP configs, agent definitions, test harnesses. If you don't sort out local exclusion early you end up with a bloated .gitignore or noisy PRs. Wrote about using .git/info/exclude for this exact problem: https://blog.stackademic.com/how-to-test-mcp-servers-and-agent-configs-without-annoying-your-team-20692917488d