Claude Opus 4.8 Dynamic Workflows: How Hundreds of Subagents Actually Work

The most significant feature launching alongside Claude Opus 4.8 isn't a benchmark improvement — it's dynamic workflows, a new capability in Claude Code that changes the scale of work one person can hand off to AI. Available in research preview for Max, Team, and Enterprise plans, dynamic workflows lets Claude plan a large task, dispatch hundreds of parallel subagents that attack the problem from independent angles, deploy adversarial agents to refute each other's findings, and iterate until the answers converge — then verify the outputs before reporting back. The flagship use case: codebase-scale migrations across hundreds of thousands of lines of code, from kickoff to merge.

For a solo developer, this changes what's possible in an afternoon. A framework migration that would take a week of manual work — updating namespaces across 200 files, running tests, fixing failures, verifying nothing broke — can now happen in a single Claude Code session. For engineering teams, it means handing off repo-wide refactors, dependency overhauls, and large-scale migrations that previously consumed weeks of senior engineer time. This is parallel agent orchestration scaled up and coordinated by the model itself, rather than by you manually managing each agent.

How Dynamic Workflows Actually Work

Traditional Claude Code operates as a single agent working through a task sequentially: read files, make changes, run tests, fix issues, repeat. This works well for focused tasks but hits limits on large-scale work where the sheer volume of files and the need to coordinate changes across a codebase overwhelm a single sequential agent. Dynamic workflows solves this by changing the architecture from one agent working sequentially to many agents working in parallel under coordination.

When you give Claude a large task, it first plans the work — breaking it into components that can be tackled independently. It then dynamically writes orchestration scripts that spin up tens to hundreds of parallel subagents in a single session. Each subagent works on its piece of the problem. Critically, Claude also deploys adversarial agents whose job is to try to refute what the other agents found — a built-in skepticism mechanism that catches errors before they propagate. The system iterates, with agents refining and challenging each other's work, until the answers converge into a single coordinated result that Claude verifies before reporting back.

The coordination happens outside the conversation, which means the plan stays on track even across multi-day execution windows. Progress is saved, and the job is resumable — if a run is interrupted, it picks up where it left off rather than starting over. This persistence is what makes long-running, large-scale work practical: you're not babysitting a single session that fails if anything interrupts it.

The Codebase Migration Use Case

Anthropic's primary example is codebase-scale migration, and it's worth understanding why this is the killer application. Large migrations — upgrading a framework version, changing a dependency across an entire codebase, refactoring a pattern repo-wide — are exactly the kind of work that's tedious, error-prone, and time-consuming for humans but follows consistent rules that can be parallelized. Updating namespaces across 200 files isn't intellectually hard; it's just a lot of repetitive work where one mistake can break the build.

With dynamic workflows, Claude Code can carry out these migrations from kickoff to merge, using the existing test suite as its bar for success. The subagents handle different parts of the codebase simultaneously, the adversarial agents catch inconsistencies, and the system verifies against your tests before declaring the migration complete. A Laravel migration that would take a week manually can compress into a single session. This connects directly to the broader trend we covered in the state of AI coding tools: AI is moving from autocomplete to agentic engineering, and dynamic workflows is the clearest example yet.

The Honest Limitations

Dynamic workflows is powerful, but it's a research preview, and there are real caveats. First, token consumption is substantial — running hundreds of parallel subagents over hours requires proportionally more compute, which means proportionally more tokens. This is expected behavior, not a bug, but it affects cost and rate limits. Anthropic raised Claude Code rate limits specifically to accommodate the higher token usage of dynamic workflows and higher effort levels, but you should still expect a large migration to consume significant resources.

Second, it's a research preview with rough edges. Anthropic's own guidance and independent reviewers both caution against pointing it at production-critical migrations without review. The "verify before reporting" step and the adversarial agents reduce errors, but they don't eliminate them. For anything where a mistake has serious consequences, human review of the output remains essential — the same principle that applies to all AI-generated code given the documented security risks of unreviewed AI code.

Third, availability is limited to Max, Team, and Enterprise plans (admin-enabled for Enterprise at launch). It's not available on lower-tier plans, and the API access model differs by plan. If you're on a Pro plan, you won't have access to dynamic workflows yet.

How to Use Dynamic Workflows

If you're on a qualifying plan, using dynamic workflows is straightforward: tell Claude Code explicitly to create a workflow for your task. Phrasing like "create a workflow for this migration" prompts Claude to plan the fan-out itself rather than working sequentially. For best results, point it at a well-tested codebase — the existing test suite is what Claude uses to verify success, so comprehensive tests mean more reliable migrations. Start with a non-critical migration to understand the behavior before trusting it with important work.

Clear instructions matter even more for large-scale agentic tasks, because ambiguity gets multiplied across hundreds of subagents. The free Prompt Optimizer helps you write precise task descriptions that reduce the chance of subagents misinterpreting the goal, and TresPrompt brings prompt optimization into your workflow. For broader Claude Code guidance, see our complete Claude Code guide.

How Dynamic Workflows Compares to Manual Parallel Agents

Developers who've worked with parallel AI agents before will recognize the core idea behind dynamic workflows, but the orchestration is fundamentally different. Previously, running multiple agents in parallel meant you — the developer — had to design the orchestration: split the task, assign work to each agent, coordinate their outputs, handle conflicts, and verify results. This worked, but it was labor-intensive and required real expertise in multi-agent design. The orchestration logic was your job. Dynamic workflows moves that orchestration into the model itself. Claude writes the orchestration scripts, decides how to fan out the work, and manages the coordination — you just describe the goal.

This shift matters because it lowers the barrier dramatically. Multi-agent orchestration was previously the domain of developers who understood agent architecture deeply. Now, the capability is available to anyone who can clearly describe a large task. The adversarial refutation step — where agents try to disprove each other's findings — is particularly valuable because it's a sophisticated quality-control mechanism that most developers wouldn't build by hand. By baking it into the workflow, Anthropic gives every user access to a self-checking system that catches errors before they reach you.

Where This Fits in the Future of Software Engineering

Dynamic workflows is a concrete signal of where AI-assisted engineering is heading: from a model that helps you write code to a system that executes large engineering operations under your direction. The winning pattern in 2026, as we've covered in our analysis of the AI coding landscape, is a control stack — project rules, reusable skills, bounded subagents, and deterministic tools around the model. Dynamic workflows is the subagent layer of that stack, productized and made accessible. It represents the maturation of agentic coding from a promising idea into a practical capability that can handle real, large-scale work.

For engineering teams, the strategic implication is worth considering carefully. Tasks that were previously too large or tedious to justify — long-deferred migrations, dependency upgrades that everyone avoids, repo-wide refactors — become feasible when one engineer can supervise an AI handling the mechanical work. This doesn't replace engineers; it redirects their time from mechanical execution to design, review, and judgment. The engineer's role shifts toward defining what should happen and verifying that it happened correctly, while the AI handles the laborious how. Used well, that's a genuine multiplier on what a small team can accomplish.

Frequently Asked Questions

What are dynamic workflows in Claude Code?

Dynamic workflows is a research-preview feature that lets Claude Code plan a large task and run hundreds of parallel subagents in a single session. The subagents attack the problem from independent angles, adversarial agents try to refute their findings, and the system iterates until answers converge before verifying outputs. The main use case is codebase-scale migrations across hundreds of thousands of lines of code.

Which plans have access to dynamic workflows?

Dynamic workflows is available for Claude Code on Max, Team, and Enterprise plans. For Enterprise, it's admin-enabled at launch. It's on by default for Max and Team plans and available via the API. Pro plans do not have access to dynamic workflows in the initial release.

Does dynamic workflows use a lot of tokens?

Yes — substantially more than a normal Claude Code session. Running hundreds of parallel subagents over hours requires proportionally more compute. Anthropic raised Claude Code rate limits to accommodate this. Expect a large migration to consume significant tokens, so factor that into your usage planning.

Can I use dynamic workflows for production code?

With caution. It's a research preview with rough edges, and both Anthropic and independent reviewers recommend reviewing outputs before merging production-critical changes. The adversarial agents and verification step reduce errors but don't eliminate them. Start with non-critical work, ensure your test suite is comprehensive, and review the results before trusting it with important migrations.

How do I start a dynamic workflow?

On a qualifying plan, tell Claude Code explicitly to create a workflow — phrasing like "create a workflow for this task" prompts Claude to plan the parallel fan-out itself. Point it at a well-tested codebase since Claude uses the existing test suite to verify success. The job is resumable, so an interrupted run will pick up where it left off.

Disclosure: Some links in this article are affiliate links. We only recommend tools we've personally tested and use regularly. See our full disclosure policy.