Anthropic shares how to make Claude code better with a harness
By FrostNinja842 · REVIEW · 3/5 stars · About Claude Code
I just read Anthropic's new blog post about harness design for Claude. The author addresses two main problems Claude faces when working for extended periods:
\- Context anxiety: loss of coherence over long periods
\- Self-evaluation bias: Claude often praises his own work even when the quality isn't good
The solution is to use multiple agents working together, drawing ideas from GANs:
\- Generator: creates code and design
\- Evaluator: provides critical evaluation and feedback
Frontend: Use 4 scoring criteria (emphasizing aesthetics and creativity) to avoid generic designs. After 5-15 revisions, the result is much more beautiful and unique
Full-stack: Use 3 agents (Planner - Generator - Evaluator)
Comparison of the same game development requirements:
\- Running alone: fast but the game has serious bugs.
\- Using a harness: more time-consuming and expensive, but significantly higher quality, beautiful interface, playable game, and added AI support.
The article also suggests that when the model becomes more powerful (like Opus 4.6), unnecessary harness elements should be removed.
Link: [https://www.anthropic.com/engineering/harness-design-long-running-apps](https://www.anthropic.com/engineering/harness-design-long-running-apps)
Anyone using Claude to code or build agents should give this a try.
636 upvotes · 6 comments
Comments (6)
PixelSurfer27: I'm real skeptical of the shovel-peddler telling me the way to fix my shovel problem is to get my shovel a bunch of tinier shovels.
NeonFlux225: I discovered this a while back. Costs a lot in tokens, but building a full workflow skill that has sub agents doing singular tasks each with their own specific job is much more effective.
Pseudo Code writers, test writers, code writers, contract definer , code reviewers, compliance reviewer agains
ThunderPulse5561: they should give more usage then, that’s the best way of making claude code good
FrostNinja842: Who hasn't been doing this? Pipelines, workflows... What do they think they are.. but I get it 'harness'.. thats the breakthrough.. let's use a different word to be original and innovative
NeonFlux225: So basically spec driven development
ShadowVibe354: **TL;DR of the discussion generated automatically after 50 comments.**
**The consensus is... complicated, but leaning heavily towards "thanks, I hate it."**
The top comment perfectly captures the mood, calling Anthropic a "shovel-peddler" for telling us the solution to our shovel problem is to buy
More discussions about Claude Code on HonestUse