Public story · 2026-06-28 · high
A grader stage forces Claude Code subagents to revise
The rubric is defined per subtask up front, and failing output goes back to the same subagent instead of straight through to the next stage.
Why now: The technique is a build note for anyone chaining Claude Code subagents together right now, not a response to any single event.
Story
A grader stage scores each subagent's output against a rubric and bounces failures back for revision, per a Totalum blog post on Claude Code subagent design. Most subagent pipelines skip this and just trust the orchestrator's judgment call on whether a subtask's output is good enough to use. That's a guess wearing a decision's clothes.
The fix is to define the rubric per sub-task, before the subagent runs, not after. When the output doesn't clear the bar, it goes back to the same subagent for another pass instead of moving downstream as-is. The orchestrator's job shifts from judging quality to routing pass/fail results.
That's the real change here: a grading gate turns quality control from an implicit hope into an explicit checkpoint you can point to when a pipeline output goes wrong. Skip the gate and every bad subtask output ships silently into whatever the orchestrator assembles next.
The advice reads less like breaking news and more like a build note, worth keeping around for anyone chaining Claude Code subagents together right now.
Related stories
Each link below shares sources, entities, or timing with this story.
Shared entity: Totalum / Same source domain / Earlier coverage
"Write once, run on any agent" becomes the skills bet.
Both cover Totalum; reported by the same outlet (totalum.app); earlier Totalum coverage from 2026-06-27.
Shared entity: Define / Shared topic / Earlier coverage
One Person, Seven AI Agents, Full Production SaaS in 14 Days
Both cover Define; overlapping topics (define, each); earlier Define coverage from 2026-05-02.
Use Vercel's json-render for structured LLM output to rendered UI.
Both cover Define; overlapping topics (define, output); earlier Define coverage from 2026-03-29.
Shared entity: Define / Earlier coverage / Tension
GPT-5.5 Is Here. It's Fast, It's Expensive, and It Changes How You Prompt.
Both cover Define; earlier Define coverage from 2026-04-25; pushes against this story (vs).
Shared topic / Tension
Chain Claude Code skills into pipelines, but let the orchestrator own control flow.
Overlapping topics (control, orchestrator, output); pushes against this story (but).
10 actionable skills from today's findings:
Overlapping topics (control, each, output); pushes against this story (vs).
Shared entity: Define / What happened next
Chain-of-Symbol prompting beats Chain-of-Thought on spatial tasks
Both cover Define; picks up the Define thread on 2026-07-02.
Shared topic
Implement Agent Skills as Auditable Quality Contracts — Intermediate
Overlapping topics (clear, control, define, explicit).
Source trail
Entities
Claim evidence
- A grader stage forces Claude Code subagents to revise
Provenance
- Canonical issue
- 2026-06-28
- AI generated
- yes
- Story unit
- 2026-06-28-wire-a-grader-revision-loop-into-your-subagents
- Labels
- source-backed, canonical briefing excerpt