Harness Design for Long-Running Application Development
Source
Anthropic Engineering
Author
Anthropic Engineering
Published
Why it matters
If you're building agents for long autonomous coding sessions, this explains two concrete fixes worth adopting: resetting context instead of compacting it to avoid degraded coherence, and running a separate evaluator agent to counter LLMs' tendency to over-praise their own generated work.
Anthropic engineer Prithvi Rajasekaran describes a three-agent harness (planner, generator, evaluator) inspired by GANs that improves Claude's performance on both subjective frontend design tasks and long, autonomous multi-hour coding sessions.
The post explains how context resets (rather than compaction) prevent agents from losing coherence or exhibiting premature 'context anxiety,' and how separating a generator agent from a dedicated evaluator agent counters LLMs' tendency to over-praise their own work.
