Anthropic's Multi-Agent Approach to Long-Running Autonomous Coding and Frontend Design

Over the past few months, I’ve been diving deep into two interconnected challenges at Anthropic: getting Claude to produce high-quality frontend designs and enabling it to build complete applications autonomously. This work builds on earlier efforts where my colleagues and I made significant strides in improving Claude’s performance through prompt engineering and harness design. However, we eventually hit a ceiling.

To break through these limitations, I drew inspiration from Generative Adversarial Networks (GANs) and designed a multi-agent system with distinct roles for each agent. This approach not only enhanced the quality of frontend designs but also improved the reliability and efficiency of long-running autonomous coding sessions.

The Multi-Agent System in Action

The core of our solution is a three-agent architecture: the planner, the generator, and the evaluator. Each agent plays a crucial role in ensuring that the final output meets high standards of quality and correctness.

• Planner: This agent breaks down complex tasks into manageable chunks, creating a detailed task list based on the product specification. The planner ensures that each step is well-defined and can be executed independently.
• Generator: The generator takes the tasks from the planner and implements them, producing code or design elements as required. It uses structured artifacts to maintain context across sessions, ensuring continuity in long-running tasks.
• Evaluator: The evaluator assesses the output of the generator against predefined criteria. For frontend designs, this involves translating subjective judgments into concrete, gradable terms. In coding tasks, it verifies the correctness and usability of the generated code.

To make this system work effectively, we had to develop a set of criteria that could turn subjective judgments like "is this design good?" into objective, measurable metrics. This was particularly challenging for frontend design, where aesthetic preferences can vary widely. We tackled this by defining specific aspects such as layout balance, color harmony, and user interface consistency.

In the context of long-running autonomous coding, we carried over two key lessons from our earlier harness work:

• Decomposing Tasks: Breaking down complex tasks into smaller, manageable chunks is essential for maintaining focus and avoiding drift. This approach ensures that each step can be executed and verified independently.
• Using Structured Artifacts: Context preservation is critical in long-running sessions. We use structured artifacts to pass context between sessions, ensuring that the agent can pick up where it left off without losing track of previous work.

Key Takeaways

1. Inspiration from GANs: The multi-agent system draws inspiration from Generative Adversarial Networks (GANs), using a generator and evaluator to improve output quality.
2. Three-Agent Architecture: A planner, generator, and evaluator work together to handle complex tasks in both frontend design and long-running coding.
3. Decomposition and Context Preservation: Breaking down tasks and using structured artifacts are crucial for maintaining focus and continuity in autonomous coding sessions.
4. Objective Criteria for Subjective Judgments: Translating subjective judgments into concrete, gradable terms is essential for effective evaluation in frontend design.

By implementing this multi-agent system, we've significantly improved Claude's performance in both frontend design and long-running autonomous coding. This approach not only enhances the quality of the final output but also increases reliability and efficiency, setting a new standard for agentic coding at Anthropic.