Where the agent's value lives is moving down the stack

The model used to be the agent. In the stacks now reaching open release, the model is just one of the layers.

That is the shift worth seeing in DeepSeek's open-source Harness, a developer preview where every component sits behind the same plugin boundary. The lab that built its reputation on open weights is publishing the connective tissue in the open, and putting the model in the same tier as the runtime chrome. On DeepSeek's product page, the architecture places models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI on the same swappable list.

When every layer sits behind one plugin boundary, the weights stop being the obvious moat. A team that can reconfigure a model, a session log, and a UI through the same harness can move between providers, runtimes, and storage backends without rewriting the agent. Selection sits in front of causation here: a public release from a major open-weight lab does not prove the value moved, but it does show where that lab is planting the flag.

Cordis, the plugin system DeepSeek credits on its product page, makes the contract concrete: the boundary is uniform, the Trajectory view sits over an append-only session log, and resume, fork, search, and replay all read the same event stream. The bet is that the agent's real product is the runtime, and the runtime is the part DeepSeek wants to open.

The honest counter: "everything is a plugin" can still be marketing framing, and adoption and performance are not stated by DeepSeek. Treat this as a data point, not a verdict. The mechanism on offer is real: when the model is just another plugin, the moat moves.