Multi-agent

Long-running AI agents can produce more code and research than we can absorb. That output extends our capacity, but it does not automatically become our knowledge, judgment, or ability.

A Claude account ban forced our fallback stack back into the main role. This is how GPT, OpenCode, iKanban, and a skill-based AI Workflow kept work moving, where they still fall behind dynamic workflows, and why orchestration should become a portable plugin rather than a runtime feature.

We used to split tasks by hand and launch agents one by one. Claude Code dynamic workflows (ultracode mode) now write their own orchestration script and run a fleet of subagents, shown through three real workflows that built our iread RSS reader, leaving asynchronous agent-to-agent communication as the next open problem.

After months of building tools and workflows on top of OpenCode, we are moving the daily driver back to Claude Code. The reasons are concrete: model quality after the GPT account-security change, ban risk when routing Claude through OpenCode, prefix-cache misses that burn the weekly quota, and a browser-wrapper experiment that did not return enough value to justify its complexity.

The next step after vibe coding is not removing humans entirely, but removing them from every inner development loop. When agents can verify more of their own work, multi-agent systems stop being blocked by one human reviewer.