Anthropic Unveils Dynamic Workflows for Claude Code

Anthropic releases Claude Opus 4.8 with dynamic workflows for Claude Code.

Anthropic has released Claude Opus 4.8 and separately introduced a dynamic workflows feature for Claude Code.

Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors.

Available today at the same price. pic.twitter.com/EufxL7T1kb

— Claude (@claudeai) May 28, 2026

The tool allows AI to write orchestration scripts independently, launching dozens or hundreds of parallel sub-agents, and verifying the work before delivering the result to the user.

It is designed for complex tasks in large codebases: security audits, bug detection, framework and programming language migrations, and project modernization.

The feature is available in a preliminary testing mode in the Claude Code command-line interface, desktop version, and VS Code extension, via API, as well as in Amazon Bedrock, Vertex AI, and Microsoft Foundry.

The mode can be initiated with a direct command to create a workflow or through ultracode. It maximizes computational efforts and allows the model to decide when to engage a multi-step scheme.

Anthropic cautioned that dynamic workflows consume significantly more tokens than a standard Claude Code session.

The model breaks down tasks into subtasks, distributes them among parallel agents, then consolidates the outputs after mutual verification and attempts to refute the solutions found.

As an example, Anthropic cited the migration of Bun from the Zig programming language to Rust. Developer Jared Sumner used dynamic workflows to generate about 750,000 lines of Rust code. The port achieved 99.8% of the existing test suite, with the process from the first commit to merging taking 11 days. However, Anthropic clarified that the version is not yet used in production.

Performance metrics of the new Opus 4.8 model:

• 69.2% in SWE-Bench Pro;
• 49.8% in Humanity’s Last Exam without tools and 57.9% with them;
• 83.4% in OSWorld-Verified;
• 1890 points in GDPval-AA;
• 53.9% in Finance Agent v2.

In Terminal-Bench 2.1, Opus 4.8 lagged behind GPT-5.5 — 74.6% versus 78.2%.

Anthropic stated that Opus 4.8 has become noticeably “more honest” in performing agent tasks: the model more frequently indicates uncertainty, less often claims unverified progress, and better detects issues in its own code before delivering the result to the user.

Back in May, Anthropic published the first report on Project Glasswing — a vulnerability search program using the Claude Mythos model.