Title. Apart from toolsets which can be ported to other harnesses, why is claude code still the best harness?What makes a harness do tasks better than others? Theres so many out there and its hard to measure how they stack up. I'd really like to ditch CC in favor of Opencode but not when its just worse even with the same models.Also know by themselves when to delegate tasks to other models or is that just a tool call?
>>109934252It's a great question and I wonder this myself sometimes. Also things like if it's actually better to use subagents vs a single agent. So much of this stuff involves observing black boxes or running experiments on some set of sample problems to try to draw conclusions about what's better at what. And frankly I think a lot of the stuff being parroted today is just cargo cult mentalityDon't have an answer for you but bumping for interest
models are trained on their own proprietary tooling, that's why claude is better in claude code.
>>109934252Unclear tool description, stupid context bloat, basically everything in Hermes is doing harness engineering wrong.
>>109934252an AI API is really braindead simple, its 1 request with an input prompt and 1 response with an output. EVERYTHING else is the harness, so most of the implementation. deciding to open and read files, dividing the work up into chunks, calling tools, appending AGENTS.md to prompts, etc
>>109936435Correct. And this training comes at a cost, which is why models have been getting worse since opus 4.6 on the anthropic side, and since 4o on the openai side, on all harness-less tasks.
