It's hard to beat domain specific code. I can avoid massive prompts and token bloat if my execution environment, tools and error feedback provide effectively the same constraints.
If I had to pick only one tool for a generic agent to use, it would definitely be ExecuteSqlQuery (or a superset like ExecuteShell). If you gave me an agent framework and this is all it could do, I'd probably be ok for quite a while. SQL can absorb the domain specific concerns quite well. Consider that tool definitions also consume tokens.
We saw what Claude Code looks like inside, and it's objectively bad-to-mediocre work, but the takeaway seemed to be 'yeah but it works and they've got crazy revenue'.
That's where we're at. The harness is kind of buggy. The LLM still wanders and cycles in it sometimes. It's a monolithic LLM herding machine. The underlying model is awesome and the harness works well enough to make it super effective.
We can do so much better but we could also do worse. It's a turbulent time. I'm not super pleased with it all the time, but it's hard to criticize in many ways. They're doing a good job under the circumstances.
I see it kind of like they're at war. If they slow down to perfect anything, they will begin to lose battles, and they will lose ground. It's a highly contentious space. The harness isn't as good as it could be under better circumstances, but it's arguably a necessary trade off Anthropic needs to make.
Based on this, are there any open source harnesses that have objectively good-to-excellent work in their code?
Do you have an example to contrast by what measure is good besides your word?
Anthropic made the most popular harness for developers.
Anthropic made the most popular desktop tool for AI automation.
And codex still uses phrases and syntax in prose ostensibly for the user as though they forgot people are actively reading this stuff.
Product is unquestionably where Anthropic excels. It is what carried it through periods where its thinking model lagged.
Excited to try it and surprised to find out about it in a random HN comment.
EDIT ADDED: bummer - looks like this only works with API keys, not Claude Pro or Max or similar for Codex
But part of me just wishes they would go back to developing and refining an excellent and user-friendly harness.
I can't imagine what the long term support is for the dozens of products they release every three months.
Meanwhile, they're shipping a more and more buggy and Byzantine Claude code with a million switches and tons of ways to use it wrong.
The subscription play really does feel like a bait and switch lock-in: "we can focus less on the harness because people with subscriptions need to use it, and focus on growth."
Interested to see if this works out for them.
This movement is brilliant for them
Of course it is and they're not hiding it. Paying 200$ a month for the equivalent of maybe 2000$ is no secret. Theyre at the frontier of the models and they need to stay there to stay relevant. Otherwise they will fall like the majority of these "AI" companies will when the bubble bursts.
