AgentSDK and Claude CLI have to load a lot of overhead to make stateless requests, so it wouldn’t be able to transfer over in the way OpenClaw it is used in OpenClaw now. OpenClaw could create a kind of adapter over the SDK, but at that point you should probably just use Claude Code.
@MatthewBerman@AnthropicAI Amodei has said in interviews they are conservative on provisioning compute so they don’t have a big gap sink the company, Seems like that conservative approach just caught up with them. Meanwhile OpenAI is canning Sora to make up the difference.
Anthropic just revoked 3rd party support. I use Anthropic OAuth for benchmarking a memory system I’m working on. Yesterday I started testing Gemma 4 26B locally on a DGX Spark as a Haiku replacement and it works great. 5.5M token benchmark runs, fully local, 24/7.
Sometimes the timing just works out.
Funny thing is that the memory system can be used with Claude Code, and needs an OAuth from it to work. Now I might have to ask for an OpenAI OAuth on your Claude plugin 😂
Been building a memory architecture for OpenClaw. I like benchmark and test focused building, but most of the big ones don’t actually test what agents really need. Had to build that as well. More soon.
Great video. A lot of these issues though are the same thing in reality, the lack of a great memory system. A competent memory system could just handle a lot of these different DBs you’re building. I think it’s one of the most underbuilt parts of OpenClaw right now and what I’ve been focusing on building.
Agentic engineering is the future, and programmers are all going to need to become architects who read through specs and not code or become irrelevant.
Engineers who are too cautious will lose to those pushing on innovation. That is the “move fast break things” ethos in a nutshell.
Those who navigate this new world and find workflows that keep quality as high as throughput will be the big winners.
One of my favorite quotes by Alvin Toffler: “The illiterate of the 21st century will not be those who cannot read and write, but those who cannot learn, unlearn, and relearn.”
My current prompt is: create 10 specialists to review this code, split them into two groups of 5. Run the first group on the code and aggregate fixes. Be very critical of fixes vs feature improvements. Fix the issues then loop. Exit the loop when you have two passes with no bugs. Then do the same for the second group. Give me a summary at the end. It works like a charm
Claude’s bug fixing has come such a long way. Two years ago I would be bashing my head against simple syntax errors, now I’m running multi agentic loops crushing dozens of bugs in 38k lines of code while generating unit tests. This is the future.
@steipete Stoked to have one of those commits in. It adds better support for memory provider plugins. Was expecting to be stuck in PR queue hell, happy this project is keeping up with the traction.