Claude Code의 'ultracode' 모드로 '심층 검색'을 시도했더니, 스스로 4단계 파이프라인에 걸쳐 약 70개의 에이전트를 가동함
Asked Claude Code for a "deep search" in ultracode mode — it spun up ~70 agents across a 4-phase pipeline on its own
핵심 요약
Claude Code의 ultracode 모드가 복잡한 작업을 스스로 워크플로우화하여 70개의 에이전트를 생성한 사례와 그 비용 효율성에 대한 분석입니다.
- 워크플로우 자동화 — Claude가 작업을 스스로 분석해 4단계 파이프라인으로 분할함
- 컨텍스트 관리 — 루프와 중간 결과를 컨텍스트 윈도우 밖으로 빼내어 효율성 확보
- 비용 문제 — 에이전트 개별 컨텍스트 설정 비용으로 인해 과도한 요금 발생 가능성
- 사전 검토 필요 — 에이전트 실행 전 예상 비용과 규모를 파악하는 기능의 중요성
Screenshot is from a single request in ultracode mode. I asked for a deep search and instead of running it inline, Claude authored a workflow: ~70 agents fanned across discovery → benchmark → enrich → verify,
each project fetched and cross-checked independently, with live progress in /workflows and an auto-ping when it finished.
What clicked for me seeing it live: ultracode doesn't just "run more agents." It moves the orchestration plan into a script — the loop and all the intermediate results stay out of the model's context window, so
only the final answer lands back in the conversation. That's why ~70 agents doesn't drown the orchestrator.
The honest tradeoff is cost. ~70 agents = ~70 context setups, not one, each paying its own overhead at your session model's rate. It paid off here because the task was genuinely too big for one window (fetching
- cross-checking every project). For a single bug fix or a few-file change, a normal session is cheaper and faster — and ultracode quietly turning every request into a workflow is the fastest way to 10x your
bill without noticing.
I put together the full cost model + when it's actually worth it here: https://avinashsangle.com/blog/claude-code-dynamic-workflows-guide
Happy to answer questions if you're weighing this for a real codebase.
