Your harness, your memory, from the LangChain blog, argues that memory is a critical part of the harness, so anyone selling harness without memory baked into it is creating a lock-in that's not visible just yet. It works today because memory is not as well-understood widely and most current interactions with LLMs are stateless. But as agent personality, personalization and long-term memory become more important, it's important that people/organizations own their memory, own their harness.

The dominant players, Anthropic, Google, OpenAI, will want to own the memory/harness - that's what you get from features like Claude Managed Agents. Not only is the model a blackbox but so is the entire persisted state that makes a model useful.

It reminds me of the same problem at the semantic layer: most cloud data warehouses, BI tools have had their own semantic layers, which is what makes analytics tick. Vendors would want it to be on their stack. LookML is a good example and is the most attractive layer of Looker. OSI, Open Semantic Interchange, is looking to change that. So you can take the semantic layer with you to any warehouse/BI tool vendor you wanna use, at least that's the promise.

But memory and agent harness are a tighter form of lock-in than the semantic layer. Semantic layer is largely static and defined by humans, updated occasionally. Memory is deeply dynamic in nature. It's seeded by you, but takes a life of its own over time.

Deep Agents is LangChain's answer - an open-source agent harness, that works with other open-source projects like LangChain and LangGraph. It's "model-agnostic," which is better than the closed ecosystems of the bigger players, but it's not as open as the author makes it sound. Deep Agents is built on Lang*(Chain/Graph) stack, which although open-source, is all owned by the same company. It's not a true interoperable solution - it's an emergent moat, where the lock-in is organically formed, rather than planned, as the agent performance is increasingly tied to whoever controls the "frameworks, runtimes, and harnesses", as Harrison himself makes a distinction of their offerings.

How it'll likely play out: You use a model provider with Deep Agents. You wanna switch models tomorrow - you can keep the harness, switch the models. Good. But your agent harness is only as good as LangChain and LangGraph, which define the primitives and the persistence/memory layer respectively. Memory also encompasses the logs generated from agent behavior, making it dependent on LangSmith, LangChain's commercial observability product. Over time, the harness works - or works better - only with Lang* ecosystem, which creates harness lock-in. You can switch model providers, but can't switch your harness ecosystem. How the agent summarizes, compacts information, what it remembers or discard - are all at the mercy of the Lang* stack.

Although it always comes with the promise of self-hosting and customization for your needs, most organizations will not or cannot do it. This has always been true for critical infrastructure, but is especially true in the LLM ecosystem given the novelty, the limited understanding most organizations have of how agents work under the hood, and the speed at which this space is evolving.

Managed solutions are likely the end-game.

So the play here seems to be to start from open-source - rather than closed-source like the dominant players - to gain market share, and convert that into structural lock-in, after significant customer adoption.