| Plain Prompting | Subagents | Dynamic Workflows | |
|---|---|---|---|
| Orchestrator | You guide Claude turn by turn | Claude spawns workers and decides next steps | A JavaScript script controls the entire operation |
| Results location | In the conversation context | In the conversation context | In script variables (out of context window) |
| Scale | Single exchange | A few workers per turn | Up to 1,000 agents per run, 16 concurrent |
| Resumable | No (turn restarts on interruption) | No (turn restarts on interruption) | Yes, within the same session |
| Best for | Simple, focused tasks | Small to medium delegation | Large scale audits, migrations, adversarial verification |
0
Maximum agents per workflow run
0
Maximum concurrent agents
0
Intermediate results in context window
Dynamic workflows are in research preview
Workflows launched on 28 May 2026 alongside Opus 4.8 and require Claude Code v2.1.154 or later. They are available on all paid plans (Pro, Max, Team, Enterprise) with Anthropic API access, and on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. On Pro plans, workflows must be enabled manually in
/config. Features and limits may change during the preview period.Start with /deep-research
The easiest way to experience workflows is the built-in
/deep-research command. It fans out web searches across several angles, fetches and cross-checks sources, has independent agents vote on each claim, and filters out anything that does not survive cross-checking. The result is a cited report with high confidence findings. This gives you a feel for how workflows operate before you start triggering them for code tasks.