<Section title=“What this term means here”>In this channel's material, agent swarms refers to Anthropic's feature for running many sub-agents in parallel — an extension of the parallel sub-agents that Opus 4.6 already spawns by default for research, code exploration, and knowledge-work tasks. The clearest evidence comes from Kieran Claassen, whose compound engineering workflow depends on parallelism: with Anthropic's agent swarms enabled, his /LFG command dropped from 20 minutes to 5.</Section><Section title=“Where swarms help — and where they don't”>Kieran's key caveat scopes the term precisely: swarms shine in implementation, where the plan and dependency graph are already clear and work can be split across agents. In planning phases, by contrast, research needs refinement rather than breadth, and parallel agents add less value.</Section><Section title=“Swarm in context”>Agent swarms sit alongside related parallelism behavior in the channel's coverage: Opus 4.6 parallelizes by default, aggressively spawning parallel sub-agents (which Naveen described as the model being “hungry for context”), while Codex only has sub-agents in experimental mode. Agent swarms are Anthropic's packaged version of that parallelism — and in the team's testing, parallelism emerged as one of the real differentiators between the two labs' coding agents.</Section>