Subagents considered harmful: lossy summaries cost more and drop context
rseroter · x · 2026-11-17
nikvdp's post "Subagents considered harmful?" argues that spawning subagents from a summary is usually a bad deal:
- In his test, the main agent's summary of a task (compare my tool to Anthropic's and OpenAI's shipped features) completely omitted the competitors, so the subagent diligently produced a useless report with no way to detect the omission.
- Citing Alexander Yue's numbers on a 137k-token Opus session: doing the task itself took the main agent 67s / $0.38, while delegating via a summary took 123s and cost more — subagents can be slower and pricier, not faster and cheaper.
- He catalogs different delegation approaches and their failure modes, why teaching an agent to choose between them was hard, and a fork-based approach that preserves richer history/context instead of a lossy summary.
A hands-on engineering retrospective highly relevant to agent orchestration work.
More from coding & agent
- After FAA ground stop, Code Metal CEO argues formal methods must modernize critical software — toddhooper · 2026-11-17
- CMU paper: planning agents should show plans, not ask — acceptance jumps from 12% to 55% — rohanpaul_ai · 2026-11-17
- Alsona's outreach agents hit 992 ICP contacts, landing #1 Product Hunt week and 103 signups — HeyToha · 2026-11-17
- Running a company in 2031: a meme of one human in a meeting with nine Claude agents — andrew_n_carr · 2026-11-17
- Perplexity to upgrade its Slack agent Computer into a persistent assignable Teammate — AravSrinivas · 2026-11-17
- One chart mapping voice AI: S2S vs cascaded, half- vs full-duplex — rdesh26 · 2026-11-17