Anthropic Explains How Claude Builds Its Own Execution Harnesses

Anthropic has published additional details about the orchestration system behind Claude Code's recently introduced Dynamic Workflows, highlighting how the feature generates custom execution harnesses designed to coordinate teams of AI agents for complex tasks.

The initial announcement emphasized Dynamic Workflows for large software engineering projects, while the new overview highlights how Claude creates and manages these workflows. Anthropic states that Claude can dynamically generate JavaScript harnesses to delegate tasks, assign agents, validate results, and determine workflow duration.

The company argues that this approach helps address several challenges associated with long-running AI tasks. Anthropic identifies issues such as "agentic laziness," where an AI system stops before fully completing a task, "self-preferential bias," where a model favors its own conclusions during evaluation, and "goal drift," where objectives become diluted over extended interactions.

Dynamic Workflows utilize multiple independent agents, each with specific roles, instead of a single context window. Anthropic outlines strategies for Claude, including "fan-out-and-synthesize," where tasks are divided into parallel subtasks and then merged, and "adversarial verification," where reviewer agents challenge the findings of other agents.

The company also highlighted tournament-style workflows, in which multiple agents attempt the same problem using different approaches and are evaluated against one another, as well as classifier systems that route tasks to different agents based on complexity or requirements.

A notable aspect of the system is model routing. Anthropic says workflows can assign different models to different stages of a task, allowing lower-cost models to handle simpler work while reserving more capable models for tasks that require deeper reasoning.

The feature has generated mixed reactions among developers. Some users view Dynamic Workflows as a potentially significant step toward more autonomous AI systems, while others question the cost-benefit tradeoff. In a Reddit discussion, one user wrote:

It will be good one day, but right now it's just a very cool way to set tokens on fire.

Others pointed to the flexibility offered by model selection:

The dynamic workflow in Claude Code allows for precise control over the specific sub-agents used at each stage. Depending on the level of complexity, you can utilize different models for different tasks.

For instance, you can use more affordable models for workflows that do not require extensive reasoning. By doing this, you can effectively optimize and reduce the overall operating costs for each workflow execution.

The discussion reflects a broader trend in AI development, where companies are increasingly focusing on orchestration frameworks, verification systems, and multi-agent coordination as a way to improve performance beyond the capabilities of individual models alone.