5x output without 5x tokens: subagents as leverage
Read the field note below to see how we apply this pattern in real Claude Code projects.
F2: 5x output without 5x tokens
Subagents did not speed us up until we redesigned the boundaries of the work itself. Parallelism without ownership does not compound; it collides. The leverage came from task design, not from spawning more agents.
What we tried
We stopped spawning subagents and started scoping tasks. For every piece of work we now ask three questions before delegation:
- Is the input well-defined? Exact files, exact branches, exact context.
- Is the output shape fixed? A table, a list, a diff, a report. Not prose.
- Is the file boundary clean? Nobody else is writing here.
If any answer is "roughly" or "mostly", the task is not ready to delegate. Tighten it first.
Where the leverage actually lives
The parent keeps two jobs that must not be delegated: edits to anything shared, and the integration step. Everything else flows in from the edges, in a shape the parent can reconcile in minutes.
What happened
Cycle time dropped and review quality went up because each delegated output arrived with a clear acceptance shape. Reviewers stopped asking "what did this one do?" and started asking "does this match the contract?". That is a faster question.
The less obvious win was on tickets. Once we scoped every task before delegating, our ticket quality improved across the board, including the work humans were doing. A habit the subagents forced us into made the team sharper in places subagents never touched.
What we learned
- Delegate bounded work, not ambiguous work. If you cannot describe the output in a sentence, the task is not ready.
- Assign one integration owner. The parent agent edits shared code; subagents do not.
- Measure leverage by lead time and rework rate, not by how many agents ran in parallel. Spawning more agents is easy; the hard part is keeping their outputs compatible.
Result
Lead time on feature branches dropped from roughly three days to under one for bounded tasks. The rework rate moved from about one in three delegated tasks needing major revision to about one in eight. The change was not about subagents per se; it was about taking "scope before delegate" seriously enough to make it a rule. Once we did, the tool got noticeably more useful.
Quick answers
What do I get from this cable?
You get a dated field note that explains how we handle this leverage-patterns workflow in real Claude Code projects.
How much time should I budget?
Typical effort is 14 min. The cable is marked intermediate.
How do I install the artifact?
This cable is guidance-only and does not ship an installable artifact.
How fresh is the guidance?
The cable is explicitly last verified on 2026-04-15, and includes source links for traceability.
Work with FRE|Nxt
We build the production AI systems we write about.
Cables are the field notes. The playbooks come from client engagements — multi-agent systems, RAG pipelines, and LLM cost cuts that ship and hold up in production. If something here maps to a problem on your roadmap, two ways in:
Audit capacity: 5 slots/month · No pitch deck · NDA on request
Your LLM cannot read your agent state
The most common architectural mistake when building LangGraph agents is assuming the LLM can see your state fields. It cannot. The LLM only sees three things…
Prompt caching kills dynamic injection. Pick one
Building a production LangGraph agent requires middleware that injects dynamic content into the system prompt every turn. Prompt caching requires that the sy…
Sub-agents are the same agent, smaller
When you add sub-agents to a production LangGraph system, the instinct is to build them as a separate agent type with its own state schema, its own graph str…