Parallel Agents Need Workflow Design, Not Hype
The real subagent skill is workflow design: decide what should fan out, what must stay serial, and what proves the work is done.
This week’s Digg discussion around Anthropic Ultracode made the pattern legible: agents are increasingly acting like smart subroutines that can scaffold, implement, and verify in parallel. The official docs reinforce the point by describing Ultracode as dynamic workflows for substantive tasks, not just a stronger reasoning setting.
Many teams will misuse this by throwing a vague task at a parallel agent system and calling the burn rate innovation. Product builders need a tighter framing. Parallel agents work when the task can be decomposed into bounded jobs with clear handoffs, shared context, and a final verifier. Without that, you pay more for confusion at higher speed.
Use a simple loop shape for one real workflow. Trigger: a scoped ticket or PR. Context: repo, acceptance criteria, and known constraints. Tools: search, edit, test, browser, or docs. Verifier: tests, screenshot diff, or policy checks. Budget: time, token, and file-change limits. Artifact: patch plus run log. Stop condition: either the checks pass or the agent escalates with a concrete blocker.
Scored evaluation set
Take a task you would normally hand to one coding agent and rewrite it as planner, worker, and verifier roles. Specify what each role can touch and what evidence it must return before the next role runs.
Open the related LLM Evaluation exercise →Keep Going