Anthropic releases Claude Code dynamic workflows to run parallel subagents for large-scale code migrations
Subagents autonomously write orchestration scripts and self-verify code changes.
#83rohan anil@_AROHAN_Should have called it Claude in bed.
Claude@claudeaiAlso new in Claude Code: dynamic workflows (research preview). For the hardest tasks, Claude makes a plan, runs hundreds of parallel subagents, and verifies its work before reporting back. Think a migration touching hundreds of files. Read more: http://claude.com/blog/introducing-dynamic-workflows-in-claude-code
@a1zhang wait is this closer to the RLM, or now finally exactly an RLM? i guess it depends on whether the user prompts and trajectories are all available in symbolic objects - i think they are
alex zhang@a1zhangEvery day we move closer to the RLM
@a1zhang so ah i guess “an RLM is just claude code” has finally become true, since claude code just became an RLM
Omar Khattab@lateinteraction@a1zhang wait is this closer to the RLM, or now finally exactly an RLM? i guess it depends on whether the user prompts and trajectories are all available in symbolic objects - i think they are
We’re also shipping dynamic workflows in Claude Code, in research preview today! This lets Claude write an orchestration plan, fan work across 100s of subagents, and double check the results before they come back. Works best in auto mode
cat@_catwuOpus 4.8 runs at high effort by default, but for the most complex or longest running jobs, change to xhigh effort via /effort for a more thorough result. We raised Claude Code rate limits to cover the extra tokens used by xhigh effort
@jarredsumner used dynamic workflows to port Bun from Zig to Rust, ~750k lines, 99.8% of the test suite passing, 11 days from first commit to merge, hundreds of agents in parallel with two reviewers on every file. Read more at http://claude.com/blog/introducing-dynamic-workflows-in-claude-code
cat@_catwuWe’re also shipping dynamic workflows in Claude Code, in research preview today! This lets Claude write an orchestration plan, fan work across 100s of subagents, and double check the results before they come back. Works best in auto mode
Every day we move closer to the RLM
