Like many, I wanted to try DeepSeek Harness in early August. But what struck me most was having the founding research paper explained to me (with AI's help): A Programming Paradigm for Spatiotemporal Composability.
Generative AI moves at a breakneck pace, and yet its trajectory keeps suggesting an analogy with geological ages—compressed onto an almost absurd timescale. A few years ago, the LLM emerged as a raw cognitive muscle (more like "right-brain": intuitive, associative). That raw force quickly required deterministic guardrails from the outside. Some got absorbed natively into the model, through training. But many others calcified externally instead—MCP tools, skills—until this hyperstructure emerged that industry calls "the harness": a deterministic architecture that channels the model's power.
The book sitting at the top of my reading list, Max Bennett's A Brief History of Intelligence, also draws a sustained parallel between the biological evolution of organisms and the evolution of artificial intelligence systems—a kind of biomimicry, conscious or not. We didn't copy the living world on purpose, but we ran into the same constraints, so we landed on the same solutions.
The paper behind DSH, I think, introduces another step of evolution—in the Darwinian sense. I won't pretend to have fully read or understood it: the math is well beyond me, and probably beyond most of us. But what I take from it is enough to think we may soon shift into a new geological age for LLMs.
In substance: the harness itself stops being a purely fixed exoskeleton. Part of it can now be composed by the model itself—not all of it, just a well-defined subset (what the paper calls "spatiotemporal composability," at the core of Cordis, DSH's microkernel plugin framework). The model can define a new component, load it into its own runtime, watch it work or take it out—without restarting the system, without touching the source code.
Something closer to a bone than an exoskeleton: still rigid, still deterministic in its logic, but integrated into the organism rather than strapped to it—capable of a slight remodeling, on the muscle's own timescale and initiative. This exists in the biological world too, an effect biologists call Wolff's law: bone grows with the organism, and in a way follows the constraints the muscle places on it.
In the current implementation, this bone regrowth doesn't survive a restart: the components the model creates live in memory, for one session. Closer to self-healing than to a persisting mutation. Still, nothing stops a human, tomorrow, from freezing into permanent configuration what the model reshaped overnight.
Maybe this is what evolution looks like here: not just a bigger muscle, but a more articulated skeleton—each piece still rigid, the whole finally able to move with what it carries.
