Subagents in practice In the captured material, subagents show up wherever a main agent decides that a task is too big or too parallel to handle alone. The clearest example is Claude Opus 4.6, which the Every team observed parallelizing by default: it spawns sub-agents for research, code exploration, and even knowledge-work tasks — running email, research, and Notion updates simultaneously. Naveen even saw it re-spawn sub-agents in follow-up turns without being asked, and described the model as 'hungry for context.' What they actually do A subagent in this material is an offshoot worker the main agent delegates a slice of a job to, then folds results back into the parent's answer. Two concrete effects stand out: - Research wins: Opus 4.6's famous one-shot fix of a two-month iOS microphone problem came from research, not raw coding genius — it synthesized a competitor's changelog and GitHub implementations through parallel sub-agents. - A consistency trade-off: the same aggressive sub-agent spawning is flagged as something that lowers Opus's consistency floor — more parallel activity, more places to drift. Not every model has them The material draws a sharp contrast: Opus parallelizes by default, while Codex only has sub-agents in experimental mode. That difference is what makes Anthropic's agent swarms such a speed lever — Kieran Claassen reported his /LFG compound-engineering loop dropped from 20 minutes to 5 minutes with swarms enabled, though he caveats that swarms shine in implementation (where the plan and dependency graph are clear) and help less in planning phases that need refinement rather than breadth. - A subagent is a delegated offshoot of a main agent, spawned to run a slice of work in parallel. - Opus 4.6 spawns them by default — for research, code exploration, even email/Notion tasks — and Naveen called the model 'hungry for context.' - Subagents powered Opus's one-shot iOS fix by gathering context from multiple sources. - They cut both ways: aggressive spawning lowers Opus's consistency floor. - Codex's sub-agents are still experimental; Anthropic's agent swarms took Kieran's /LFG loop from 20 minutes to 5.