The AI coding stack just split in two

Bun is writing docs for AI agents, Ollama is teaching you to run Claude Code on a local model, and the coding-tool fight has quietly moved up the stack.

The week's signals reveal a quieter but more consequential shift in AI coding tools: the model layer is commoditizing while the workflows around coding agents are becoming first-class engineering artifacts. When Bun's experimental Rust branch ships with a PORTING.md guide written for coding agents and Ollama publishes Claude Code integration docs showing how to swap a local GLM-5 in for Anthropic's hosted model, the fight has moved up the stack.

Three pieces of evidence make the case. First, Bun maintainers — exploring an experimental Zig-to-Rust port — wrote documentation explicitly targeted at coding agents, not human contributors, treating the agent as the primary workforce on a multi-month migration. As Simon Willison noted on May 5, 2026, this is an experimental branch rather than a confirmed switch, but the artifact itself — a PORTING.md aimed at agents — is the data point.

Second, the model substrate underneath those agents is becoming generic. The Ollama docs (May 2026) demonstrate Claude Code running against a local GLM-5 endpoint with a 64k+ context window, meaning the harness Anthropic ships is increasingly decoupled from the model that powers it. AWS shipped four new Qwen models in SageMaker JumpStart on May 4, 2026, with explicit agentic-coding framing on Qwen3.5-27B-FP8 and Qwen3.6-35B-A3B.

Third, vendors are picking opposing positions on autonomy. Mistral's Leanstral, per The New Stack, wants to remove human-in-the-loop code checks. Anthropic's new research institute TAI, announced May 7, 2026, explicitly lists "human visibility into and control over" self-improving AI systems as a core research agenda. Same week, same problem space, opposite philosophies.

What this means

For builders: stop optimizing for a specific model. The diyai.io listicle crowning Claude Code best overall and Cursor best IDE for 2026 is already date-rotting — by next quarter the harness-versus-model split will matter more than the brand on the box. Start writing PORTING.md-style guides for your own repos.

For buyers: the procurement question shifts from "which model is best at coding" to "which agent harness lets us swap models when prices move." Lock-in risk now lives at the agent layer, not the model layer.

For creators: Simon Willison's observation that vibe coding and agentic engineering are blurring matters because the working definition of engineering is being rewritten in real time.

The bottom line: Treat your repo's documentation as agent infrastructure, not just human onboarding — the projects writing PORTING.md files today are the ones that will move fastest tomorrow.

Sources