AINEWS
A former Kimi CLI lead says the engineering wrapped around AI models won't disappear as models improve: it will move up the stack, from patching capability gaps to coordinating between agents.
As AI models grow more capable, the engineering wrapped around them is not shrinking. It is migrating upward, from patching capability gaps into coordinating between agents and across sessions. That is the argument a former Kimi CLI lead makes in a long essay published Monday, a practitioner-level pushback against a popular AI narrative that says the surrounding engineering will simply be trained away.
The wrapper has a name in the trade. Practitioners call it the "harness": the runtime layer of prompts, tool calls, agent loops, and state management that lets a model actually do work rather than just answer. Most AI coverage treats that layer as scaffolding that disappears as the model improves. The opposite, the source argues, is closer to what the engineering actually looks like.
"Anyone who says the harness will be trained into the model has never built one," wrote stdrc on X on August 9, two days before the long-form essay went up. stdrc runs Raft, a startup building the same layer of coordination for teams of AI agents.
His argument has two parts. The first is the size claim: the harness is not thinning out, it is thickening. The second is the location claim: the complexity is not vanishing, it is moving. What used to be patches for missing model capability, including hand-written prompts, duct-taped tool integrations, and brittle loops that broke when the model misbehaved, is being replaced by a different kind of work: scheduling, state management, governance, and the choreography of multiple agents handing tasks to each other.
The four elements of any harness, stdrc writes, are the agent execution loop, context and state management, tool and resource scheduling, and safety and boundary governance. As base models improve, the lower rungs of that stack get simpler. The upper rungs get harder. He uses a yin-yang image to describe the trade: the lower harness "internalizes" into the model (yin expanding), while the upper harness "grows" as the coordination problem expands (yang extending). The total surface area does not shrink. It relocates.
The Kimi CLI codebase, as he describes it, is the worked example. At one point, the team planned to remove a dedicated subagent scheduler and let the model split work itself by issuing commands through a bash terminal. They planned to strip out the native parallel tool-calling layer and let the model generate a tool-call script that ran in parallel on its own. Both moves were meant to make the harness thinner. Both, in practice, pushed complexity somewhere else: into the terminal, into the script, into the runtime that had to interpret what the model produced. The harness did not vanish. It moved.
In a second X post from earlier in the year, stdrc traced how the product itself evolved: single tasks, then multi-agent collaboration, then cross-session state synchronization, then handoffs between agents. Each step added harness, not subtracted from it. The multi-agent harness, he writes, is the predecessor of what Raft is now building.
The foil in the essay is Pi, the personal-AI product from Inflection that runs on a system prompt under 1,000 tokens and four basic tools. Pi represents the maximalist version of the model-eats-everything thesis: shrink the wrapper, trust the model, keep the surface area small. stdrc's point is not that Pi is wrong as a product choice. It is that the bet generalizes poorly. Most of the work he does, and most of the work he sees other teams paying for, sits in the coordination layer Pi explicitly leaves out.
There is a falsifier in the argument, and stdrc is honest about it. If harnesses were genuinely shrinking as models improved, the practitioner evidence would look like the opposite of what he describes. It does not. The teams shipping agent products at scale, in his telling, are not seeing their coordination code shorten. They are watching it lengthen and move up the stack.
The piece does not predict a winner. It draws a line: the conversation about AI tends to collapse "the model" and "the engineering around the model" into a single thing, and to assume the engineering will be eaten. The actual practice, as stdrc describes it from inside the Kimi CLI codebase and now from Raft, is the opposite, with the harness thickening rather than thinning, and the thickening happening in the places that decide whether an agent system is reliable in production. The next test will be whether that pattern holds as base models cross whatever capability threshold the model-maximalists are waiting on next.