AIMartin Fowlerabout 2 hours ago
Harness Boosts Reliability, Cuts Cost by 33%
15 min read
Martin Fowler reports that a behavioral harness allows a cost-optimized coding model, GPT-6 Luna, to achieve 100% reliability matching a frontier model. While the harness increased runtime and token usage, it reduced the cost per run to $0.273 from $0.406, a 33% saving. The experiment tested agents building an API service where unaided runs failed in 7 of 10 attempts.
Level
Hype check
Tap to vote and see what everyone thinks.