Dynamic workflows (research preview) let Claude write an orchestration script on the fly and spin up a large fleet of coordinated subagents in parallel to tackle complex tasks. Say "workflow" in a prompt to trigger it.
Mentioning "workflow" in a prompt triggers Claude to generate a structured orchestration plan it then strictly follows, ensuring each stage completes in the correct order. The feature is framed as the most significant new addition to Claude Code.
The standout new capability in this release is dynamic workflows: Claude now plans a task, fans it out to tens or hundreds of parallel subagents, verifies their outputs, and iterates until everything converges into a single coordinated answer.
Dynamic workflows are available today across Max, Team, Enterprise, and the API (including Bedrock, Vertex AI, and Foundry). On by default for Max and Team plans; Enterprise admins enable it via managed settings.
Opus 4.8 is live and beats 4.7 on SWE-bench Pro (64.3 to 69.2). The bigger behavioral shift is honesty: it flags uncertainty, catches its own bugs, and stops short of declaring victory when the work isn't done. Priced the same as 4.7.
Opus 4.8 defaults to high effort mode, spending roughly the same tokens as 4.7's default while performing better on coding tasks. For harder problems or long async work, xhigh mode is available, and rate limits have been raised to accommodate the extra token spend.
Opus 4.8 is now live and positioned as the recommended model for everyday Claude Code use. The headline behavioral change is improved honesty: it flags problems in its own code and admits uncertainty rather than papering over issues.
A dedicated prompting guide for Opus 4.8 is now live, covering tips and best practices for getting the most out of the model.
The model flag uses a hyphen, not a decimal point. The correct syntax is --model=claude-opus-4-8, and the dot variant (claude-opus-4.8) does not work.
Jarred Sumner used dynamic workflows to port roughly 750k lines of Bun from Zig to Rust in 11 days, reaching 99.8% test suite pass rate, with hundreds of agents running in parallel and two reviewers assigned to every file.
Dynamic workflows combined with adversarial code review were the key techniques behind rewriting Bun in Rust in six days, offered as a concrete proof point for the approach.
A reliable three-phase loop for each unit of work: do the work (no slow commands), run adversarial review, then apply changes. Works for bulk rewrites like porting every .zig file to Rust or getting a whole crate tree to compile.
A concrete workflow pattern: step one fixes a bug from a report file, step two spins up two adversarial review agents to poke holes in the fix. Git and build commands are banned per step to avoid conflicts when multiple Claude instances share a branch.
Each workflow step shares the same prompt template but runs in a forked context window, making the process more deterministic than open-ended subagent delegation. The mental model is closer to a custom build pipeline than a chat session.
Giving each workflow step its own context window keeps Claude focused on exactly the task at hand, which empirically improves results. Individual workflows have been observed running for 10 hours continuously.
You can tell Claude roughly how many tokens to spend, nudging it toward thoroughness on hard tasks and frugality on easy ones. This works especially well inside dynamic workflows.
Salesforce published a detailed writeup on adopting Claude Code at scale. A migration originally scoped at 231 days shipped in 13, and one PR delivered 21 endpoints at 100% test coverage.
Despite a large increase in shipping velocity, total incidents dropped 5%. Security guardrails and quality standards were built directly into the agentic workflow, which challenges the usual assumption that speed and quality trade off against each other.
