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

Anthropic’s article examines harness design for long-running AI agents, showing practical experiments building a 2D game and a browser DAW. It analyzes failure modes like context anxiety and poor self-evaluation, and proposes solutions: planners, generator/evaluator loops, context strategies, and adversarial evaluation. The post clarifies how harness components must evolve as models improve.

– What a harness is: the orchestration layer around a model (prompts, tools, feedback loops, constraints, validation) that keeps agents on track for multi-hour or multi-day tasks.

– Key failure modes: context anxiety (models rush to finish as context fills) and poor self-evaluation (agents overrate their own output); mitigations include context resets/compaction and dedicated evaluator agents.

– Architectures and experiments: patterns shown include initializer/planner, generator, and adversarial evaluator; sprinted handoffs versus continuous 1M-token runs (Opus 4.6) were compared using a 2D game and a browser DAW.

– Practical takeaways: evaluator design needs graded, weighted criteria plus interactive testing tools; harness assumptions must evolve with model capability — simpler harnesses may suffice as models improve.

Quotes:

A wild horse has raw power, but the harness allows you to control the power, set it in a direction and get where you want to go.

As the context window fills up, models don’t just lose coherence – they start wrapping up the conversation prematurely.

If you ask an agent to evaluate its own work, it’s likely going to praise it.