Agent harness is the future of product #claudecode #openclaw #aiagent #harnessengineering #claudecodeleak

Agent Harness is the future of product. For anyone who views himself as a product manager, I'm sure you probably have some flavor of this self-question at some point over the last two years, which is that assuming the model just continues to get better, is there even a point for product or user experience design anymore? Will the model just be the one and only all encompassing product? And if so, do we even need PMs in the future? But on the other hand, some of my favorite agentic products just feel so different from one another, despite having similar capabilities. For example, even when using the same online model, clock code and open-claw just feel so different. The conclusion that have reached, at least till now, is that models are not everything, and their agentic harnesses are the future of product. So to show what I mean and thanks to the leak of clock code source code earlier this week, I went ahead and compared clock code and open-claw, and I wanted to share with you why subtle design differences in their harnesses can lead to fundamentally distinct product experiences. So first, despite both being general agent exhibitions, they're designed for different core experiences. Clock code, as its name suggests, is first and foremost created as a developer-first agent, and because us developers work in a context of a co-base, it thus has what I call a project or repo-based experience. On the other hand, for open-claw, it's created to be like Tony Stark's Jarvis, a persistent, always-on AI assistant, so its focus is what I call an agent-based experience. Let's unpack a bit of what I mean here. A great example of how the two agents differed in order to best serve their respective core experiences is your memory and context design. Clock code's memory is largely project or repo-scoped. If you open up clock code in different projects, you will have completely different memories. On the other hand, open-claw's memory is agent-scoped. Each agent has a workspace directory that contains all of its memory files. Instead of tracking memories by project, open-claw records memory by day, and they're all accessible by the agent, whereas clock code, on the other hand, can only access memories within a particular project. Session designs are also fundamentally different between the two agents. Clock code, because it's built initially as a developer tool, is primarily optimized for a single user, single session experience in a terminal. Each session is user-driven, meaning that the user controls when to open the new clock code session and its context is fully isolated from that of the other sessions. Open-claw, on the other hand, is designed for cross sessions at multi-channels from the start. The concept of a session, in fact, is almost hidden away from the end user. Contacts across channels are unified into sessions by default, which is why you feel like talking to the same person all the time. Clock code and open-claw each further optimizes its core experience in different ways. Clock code, as its developer centric, provides code, diff, UI, plan mode, and explicit user controls over context management. Open-claw, on the other hand, almost once its user to not even have any awareness of context limit at all. And anyways, around things like soul.md, and makes the agent feel more personalized. All in all, these are all subtle harness designs that facilitate towards the core agent experience its creators want to manifest. So yeah, I think product and UX design, in my opinion, are going to be different, but they're still tremendously important in the age of AI agents. At least I'm not convinced otherwise yet.