7 posts on Agent Orchestration.
A production agent orchestration pipeline has eight stages: trigger, plan, route, isolated execution, approval gates, deliverables, memory and resume. How each works, which multi-agent architecture to pick, what breaks in practice, and how Kavela runs it.
What Claude Code orchestration means in practice: subagents, parallel sessions in git worktrees, and a lead agent that splits the work. Where it breaks (handoffs, ownership, approvals, sessions that die) and how to run it with a manager agent on your own Claude plan.
Yes, Claude can manage other agents. How the manager and worker pattern works, how to prompt a lead agent to delegate well, the three ways it fails (the manager does the work, context gets lost, nobody reviews) and how to run workers on your own Claude plan.
How to orchestrate Codex agents in parallel: split the work, give each agent its own worktree or cloud task, track owners, collect diffs, route approvals and keep runs alive. Plus how Kavela runs Codex workers on your own Codex plan.
Running several AI agents brings file collisions, orphaned tasks, duplicate work, lost context and silent waits. How to coordinate multiple AI agents: one coordinator, one owner per task, isolated workspaces, shared context, one approvals queue and a visible schedule.
What a tool needs to orchestrate AI agents: one view of all work, clear ownership, schedules, approvals, logs and deliverables, and agents that keep running with your editor closed. How an editor, terminal multiplexing and a calendar workspace compare.
Managing many Claude Code or Codex sessions means tab sprawl, runs that die when the laptop sleeps, lost context and constant checking. Practical fixes (names, worktrees, tmux, notes, notifications) and the structural one: sessions as tasks on a calendar.