Problem
For months, we’ve treated agents as individual workers: one chat, one task, one list of steps, one execution. That works for small fixes, bounded research, or linear automations.
But many real tasks aren’t linear. Auditing a large codebase, migrating a framework, searching for technical debt, or validating security requires dividing, parallelizing, contrasting, and recomposing. Until now, humans did that architecture: deciding subprocesses, assigning reviewers, setting boundaries, and gathering results.
With Dynamic Workflows, Anthropic pushes in another direction: the agent starts designing its own operation.
Thesis
Dynamic Workflows matters less for “more subagents” and more for the role change.
Claude Code is no longer limited to executing a single instruction. It can write orchestration scripts, launch dozens or hundreds of subagents in parallel, verify findings, and return more closed work. This turns the agent into a small temporary organization: it plans, distributes, executes, and reviews.
The advantage won’t be pressing a bigger button. The advantage will be knowing which jobs warrant that deployment.
Framework
To decide if a dynamic workflow makes sense, use four tests:
- Breadth: the task can be divided into independent fronts.
- Verifiability: each front can produce evidence, not just opinion.
- Recomposition: there’s a clear way to gather findings into a decision.
- Error cost: it’s worth paying more compute to reduce human error or blindness.
Mini-case: “find dead code” in a monorepo can benefit from parallelism. Each subagent inspects an area, cross-checks imports, searches for tests, and proposes candidates. “Change this text on a page” doesn’t need it. If you use a fleet of agents for a linear task, you don’t have automation: you have dramatized expense.
Measurable signal: percentage of verified findings over findings proposed by subagents.
Posture: the future of agents isn’t just autonomy; it’s operational topology.
Why it matters now
Anthropic introduced Dynamic Workflows for Claude Code on May 28, 2026. In their announcement, they position it for complex, end-to-end tasks with orchestration scripts and parallel subagents. They also warn it can consume many more tokens than a typical Claude Code session.
The initial availability includes Claude Code CLI, Desktop, VS Code extension, API, and environments like Amazon Bedrock, Vertex AI, and Microsoft Foundry, with nuances by plan and administration. This suggests it’s not a lab demo: it’s an operational piece for teams that already want to put agents into real repos and workflows.
The practical consequence is clear: if the agent can create the workflow, the team must govern the perimeter, stopping criteria, and economy of each execution.
Anti-example
“Use Dynamic Workflows for anything big.”
Big doesn’t mean parallelizable. Some tasks are big because they’re ambiguous, political, or depend on a human decision. A dynamic workflow can multiply errors if it distributes a bad premise among many subagents.
Protocol (3 steps)
- Write the stopping condition. Before launching subagents, define when they should stop.
- Ask for evidence per finding. Each conclusion must bring a file, line, test, log, source, or reproduction.
- Reserve an adversarial reviewer. One agent must try to break the solution before it reaches a human.
| Task type | Good use of Dynamic Workflows | Bad use |
|---|---|---|
| code audit | parallel search with verification | subjective list without evidence |
| large migration | divide by modules and test | touch everything without boundaries |
| security | review repeatable patterns | decide risk policy |
| product | stress-test options | replace strategic criteria |
Related
- Claude Managed Agents + Cloudflare: the perimeter becomes a product
- Agent Reliability Score: how to know if an agent deserves autonomy
- Rollback Design for AI Workflows: how to shut down automations without breaking operation
Sources consulted
- Anthropic: Introducing dynamic workflows in Claude Code
- Anthropic: Claude Code subagents
- Anthropic: How we built our multi-agent research system
Next step
Before trying Dynamic Workflows in production, choose a broad, verifiable, and reversible task. If you can’t define evidence and stopping criteria, you don’t have a workflow yet; you have a bet.
Translated from the Spanish original with AI assistance and reviewed for accuracy. Read the original in Spanish.