When you say “go build an auth system,” the agent must research alternatives, evaluate options, and then implement — but by implementation time, its context is polluted with details about approaches it won’t use. The fix: create a research task to evaluate implementation possibilities, decide on the approach (yourself or via a decision agent), then hand the chosen approach to a fresh-context agent for implementation. The implementer gets “implement JWT authentication with bcrypt-12 password hashing, refresh token rotation with 7-day expiry” — no research residue.
This is a workflow-level application of Context inefficiency compounds three penalties: cost, latency, and quality degradation — research tokens don’t just waste space, they actively degrade implementation quality by introducing confusion. The pattern also connects to Spec files are external memory that survives context resets, where the research output becomes a spec that persists across the context boundary between research and implementation sessions.