Subagents vs skills: context isolation for agentic coding workflows¶
Insight: Subagents solve the context pollution problem in agentic coding: exploration and verification generate the most output, burying the main conversation in noise. Use skills when you want your current agent to follow a better procedure; use subagents when the work should happen in its own context and only the result matters. Subagents offer clean context isolation, parallelism, tool scoping, and model mixing. HumanLayer's testing confirms that expanding to maximum context windows (1M tokens) degraded instruction adherence compared to focused, smaller contexts — models struggle to distinguish critical instructions from noise at maximum context size.
Detail: Builder.io implements subagents via markdown files in .builder/agents/ with name, description, and tools list. The key insight is that subagents turn a single conversation into a coordinated team — each subagent gets a fresh context window for noisy work (search results, test output, exploration) and returns only a summary. This prevents the "scrolling through 400 lines of logs" problem that plagues long agentic sessions. HumanLayer's architecture recommendation: use sub-agents with "context firewalls" rather than context aggregation — delegating focused tasks to specialized sub-agents with smaller context windows prevents irrelevant information accumulation.
Sources
Related: existing entry "Claude Code tips and workflows" in external/claude-code.md — COMPLEMENTS