2026-05-30

Claude Code’s Dynamic Workflows Aim to Make Multi-Agent Coding Practical

Claude’s new dynamic workflows feel like Anthropic trying to push Claude Code from “helpful coding assistant” into “orchestrator for serious engineering work.” For developers, the interesting part isn’t the marketing language about speed — it’s the claim that Claude can now plan, fan out, verify, and keep state across large jobs instead of choking on them in a single pass.

Key Points

Anthropic has introduced dynamic workflows in Claude Code as a research preview.
The feature is meant for hard, end-to-end tasks that are too large for a single agent pass: repo-wide bug hunts, large migrations, modernization work, and heavy review tasks.
Claude can dynamically write orchestration scripts and run tens to hundreds of parallel subagents in one session.
The system includes checking and verification before results are returned, including independent attempts and agents that try to refute findings.
Dynamic workflows are available in:
- Claude Code CLI
- Claude Code Desktop
- VS Code extension
- Claude API
- Amazon Bedrock
- Vertex AI
- Microsoft Foundry
Availability is described for Max, Team, and Enterprise plans, with Enterprise availability controlled by admins.
Anthropic warns that workflows can use substantially more tokens than a normal Claude Code session.
Users are encouraged to start with a scoped task and to use auto mode for the best experience.
There are two main ways to start:
- Ask Claude to create a workflow
- Turn on ultracode, a Claude Code-specific setting that sets effort to xhigh and lets Claude decide when to use a workflow
Progress is saved during execution, so interrupted jobs can resume instead of starting over.
The coordination happens outside the conversation, which Anthropic says helps keep the plan on track during long runs.
The company highlights real-world use cases like:
- codebase-wide security and optimization audits
- detection of dead code and cleanup opportunities
- large framework, API, and language migrations
- “critical work you need checked twice”
Anthropic quotes engineers saying the system helps with discovery, review, and trust in longer runs.
The blog uses the Bun rewrite as a showcase:
- a port from Zig to Rust
- 99.8% test suite pass rate
- about 750,000 lines of Rust
- eleven days from first commit to merge
In that example, one workflow handled lifetime mapping, another wrote files in parallel with two reviewers per file, and another fix loop drove the build/test suite to green.

My Take

What strikes me is that Anthropic is not really selling “another agent.” It’s selling a whole operating model for agentic work: plan, fan out, verify, retry, converge, repeat. I think that’s the right direction if you want Claude Code to be useful on the kinds of tasks humans actually dread — large migrations, repo-wide cleanup, and finding bugs that static analysis misses.

The part I’m most interested in is the built-in skepticism. Independent verification and adversarial agents that try to break the result sound much more valuable than just spawning more subagents for the sake of parallelism. In practice, that’s what makes an agent workflow feel trustworthy instead of merely impressive.

That said, the token warning is not a footnote — it’s the main caution. I’d be curious whether teams will actually use this for everyday work, or whether it becomes a “break glass for huge jobs” feature because the cost and complexity are too high for routine tasks. My guess is the sweet spot is exactly what Anthropic says: scoped, high-value work where human review would otherwise take forever.

The Bun rewrite example is flashy, and maybe a little too perfect as a showcase, but it does make the feature feel concrete. Still, I’d want to see how this holds up on messy, real enterprise codebases with stale tests, weird architecture, and lots of context missing. That’s where the hype will either turn into something real or run into the usual agent limits.

Overall, this feels like a meaningful step for Claude Code: less chatty copilot, more coordinated engineering system. If it works as advertised, it could be one of the more genuinely useful agent features I’ve seen from Anthropic so far.