Anthropic Just Dropped the New Blueprint for Long-Running AI Agents.

Key Concepts

• Agent Harness: The orchestration layer (software, prompts, tools, feedback loops, constraints) that wraps an AI model to ensure reliable, goal-oriented execution.
• Context Anxiety: A failure mode where LLMs, as their context window fills, prematurely terminate tasks, rush steps, or falsely declare work as "done."
• Context Compaction: A technique to summarize conversational history to preserve space for active, usable context.
• Context Reset: A strategy of starting with a fresh context window, reading from a progress file, and handing off tasks to prevent context-related degradation.
• Adversarial Evaluation: A multi-agent framework (Generator vs. Evaluator) where a dedicated QA agent critiques the work of a generator agent to improve output quality.
• Ralph Wiggum Loop: An agent loop that uses external, objective validation (e.g., linters, type checkers) to ensure tasks are completed correctly.
• AI Slop: Generic, low-quality, or repetitive AI-generated content that lacks originality or craft.

---

1. The Role and Design of Agent Harnesses

An agent harness is the structural "car" that allows the AI "engine" to reach a destination. Without a harness, agents often lack direction, fail to finish complex tasks, or lose coherence. Anthropic’s research highlights that for long-running tasks (e.g., coding, compliance audits, risk analysis), the harness design is as critical as the model itself.

2. Failure Modes in Long-Running Agents

Anthropic identified two primary failure modes that occur when agents operate without robust oversight:

• Context Anxiety: Models struggle with long-term memory, leading to premature task completion. While context compaction helps, it is not always sufficient for older models like Sonnet 4.5.
• Poor Self-Evaluation: Agents tend to be overly optimistic about their own work. They often overlook subtle bugs or produce "bland" outputs, failing to maintain high standards of craft or originality.

3. Methodologies for Reliable Execution

To mitigate these failures, the following frameworks are recommended:

• Decomposition & Handoff: Breaking large projects into smaller features, committing progress to a file, and using a "clean slate" approach (context reset) for each new task.
• Adversarial Evaluation (Generator-Evaluator): Implementing a two-agent system where the Generator creates content and the Evaluator (acting as a skeptic) provides critical feedback.
  • Requirement 1: Make subjective quality gradable by defining specific principles (e.g., design quality, originality, craft, functionality).
  • Requirement 2: Weight criteria based on model capabilities.
  • Requirement 3: Provide the evaluator with tools (e.g., Playwright MCP) to interact with the output as a real user would.

4. Case Studies and Experiments

• Front-End Design (Dutch Art Museum): By using an adversarial harness, the system moved from generic designs to a unique 3D room concept after 10 rounds of feedback.
• Full-Stack Coding (2D Retro Game Engine):
  • Solo Harness: Produced a non-functional, superficial result.
  • Full Harness (Planner + Generator + Evaluator): Successfully built a functional game by splitting the project into sprints, defining "done" criteria upfront, and using contract negotiations between agents.
• Digital Audio Workstation (DAW): Using Opus 4.6, the team simplified the architecture by removing context resets and sprints, relying on context compaction and a final evaluation phase. The project took ~4 hours and cost $125.

5. Notable Quotes and Perspectives

• On Harness Evolution: "Every component in a harness essentially encodes an assumption that the model can't actually carry out that task itself."
• On Model Improvement: As models like Opus 4.6 evolve, they may require less complex harnesses (e.g., removing context resets), but the need for adversarial evaluation remains vital when pushing models to their limits.
• On QA Limitations: Anthropic admitted that "out of the box, Claude is a poor QA agent," noting that it often identified issues but then "talked itself into deciding they weren't a big deal."

6. Synthesis and Conclusion

The core takeaway is that building reliable, long-running agents is an iterative process of harness engineering. Developers must move away from "one-shot" prompting toward structured systems that include planning, adversarial evaluation, and objective validation. While newer models with larger context windows (like Opus 4.6) reduce the need for complex context management, the "Evaluator" agent remains a necessary component for ensuring high-quality, non-generic outputs. The most effective systems are those that are simple, modular, and capable of interacting with their own outputs through tools.