Related: sources · notes · metadata · Published Pieces

Apple Hardware, Weak Software

Better models do not make runtime engineering less important. They increase the return to runtime engineering.

The most charitable version of the model-lab strategy is Apple.

Apple’s control is annoying, but it is not arbitrary. Apple earns some of its vertical integration because the integration works. Hardware, software, security, battery, chips, displays, developer frameworks, distribution, and user experience cohere. The company’s control has a product argument behind it. You may dislike the walled garden, but the garden is real.

Anthropic-like labs want a similar position in AI. They do not merely want to sell model access. They want models, tools, memory, sandboxes, skills, managed agents, enterprise workflows, and eventually the whole operational layer. The pitch is integration: the model works best when the harness is designed around it.

This could be compelling if the software matched the hardware.

The models are excellent. The runtime taste is not.

The current agentic stack still feels caught between chat, files, tools, sessions, sandboxes, and platform-managed workflows. Claude Code is strong because the codebase is already a good substrate for agents: files, Git, tests, diffs, logs, compilers, CI, terminals. The model can observe the world, make changes, run feedback loops, and be corrected by reality.

Files are useful. They are not a theory of computing.

A filesystem is an affordance. It is legible to humans and models. It supports reading, writing, grepping, diffing, patching, and mounting. For coding work, files are natural because codebases are file-native. But for long-lived multiagent systems, workflows, public memory, provenance, and agentic media, files alone are insufficient. They are projections, not canonical state.

The deeper substrate should include events, actors, capabilities, permissions, durable objects, databases, traces, verifiers, reducers, provenance, and rollback. Files may be the workspace view, the import/export surface, or the temporary scratchpad. They should not be mistaken for the operating system.

This is the recurring weakness in model-lab software. The model is powerful enough to make crude substrates look smarter than they are. The lab then risks confusing model brilliance for runtime correctness.

The “harness will wither away” idea is part of the same mistake. Better models reduce the need for brittle prompt scaffolding, but they do not eliminate runtime engineering. Better models increase the return to runtime engineering.

A better harness gives a better model richer state, clearer tools, denser feedback, more reliable memory, safer action spaces, cheaper verification, and lower entropy. A better model makes better use of that harness. These are complements, not substitutes.

The lab wants to be Apple, but Apple’s lesson is not merely “vertical integration wins.” Apple’s lesson is that integration must be beautiful, coherent, and deeply designed. Otherwise vertical integration becomes lock-in without grace.

Right now, model labs have frontier hardware and immature software metaphysics. They have brilliant models wrapped in interfaces that still overcenter chat, files, sessions, and managed platform assumptions. They are moving in the right direction, but they are not yet building the final computer.

The final computer will not be a chat session with tools. It will not be a managed folder with subagents. It will not be a lab-owned Slackbot with memory.

It will be artifact-native, event-driven, multi-model, provenance-bearing, permissioned, inspectable, rollbackable, and composable. It will treat files as useful projections. It will treat chat as ingress. It will treat agents as internal organs. It will treat durable artifacts as the interface.

The model labs may get there. But they should not be trusted merely because their models are strong.

Apple hardware with weak software is not Apple.

It is a warning.