Skills2026-06-24 · source-backed
Keep concurrent subagents to 3-5 unless the task genuinely fans out.
Story
Past five, you spend more time reconciling summaries than you save fanning out. High parallelism (tens or hundreds) only pays off for work that actually decomposes, like an 80-combo benchmark sweep. Treat it as a special case you justify, not a default you reach for (source).
Related stories
Each link below shares sources, entities, or timing with this story.
Shared entities / What happened next
A wave of local-first, bring-your-own-model coding agents is forming under the cloud harnesses
Both cover Keep, Treat; picks up the Keep thread on 2026-06-29.
Shared entities / Earlier coverage
MiniMax M3: an open-weight coding model just cleared the frontier bar, and the weights are coming
Both cover Keep, Treat; earlier Keep coverage from 2026-06-08.
Prompt Engineering
Both cover HIGH, Keep; earlier HIGH coverage from 2026-02-21.
Shared entity: Treat / What happened next / Tension
Treat prompts as code with PromptFoo in CI.
Both cover Treat; picks up the Treat thread on 2026-06-28; pushes against this story (against).
Shared entity: Treat / Earlier coverage / Tension
Every published prompt-injection defense gets bypassed above 90% under adaptive attack.
Both cover Treat; earlier Treat coverage from 2026-06-23; pushes against this story (against).
Your model is an upstream dependency you don't control. Plan accordingly.
Both cover Treat; earlier Treat coverage from 2026-06-14; pushes against this story (against).
Shared entity: Keep / Earlier coverage / Tension
Kimi K2.6 Beats GPT-5.4 on SWE-Bench Pro. It's Open-Weight. It Costs 88% Less.
Both cover Keep; earlier Keep coverage from 2026-05-05; pushes against this story (vs).
Shared entity: Treat / Earlier coverage / Tension
Karpathy at Sequoia: The Spec Is the New Code, and AI Output Is 'Awkward and Gross'
Both cover Treat; earlier Treat coverage from 2026-04-30; pushes against this story (but).