A common observation about coding agents is that their capability is the product of LLM × harness. The LLM provides intelligence. The harness provides everything around it—system prompts, workflows, memories, examples, MCP servers, retrieval, guardrails, tool orchestration. Early on, the harness mattered enormously because the models weren’t yet good enough. Much of Claude Code’s early advantage came from its harness as much as the underlying model.

But this balance is shifting.

When Anthropic released Opus 5, they shared an interesting result: they deleted roughly 80% of their system prompts with essentially no measurable degradation on coding evaluations. In many cases, performance even improved.

This wasn’t because they found better prompts. It was because they found fewer prompts.

Years of prompt engineering had accumulated layers of defensive instructions, examples, exceptions, and repeated guidance—much of it written for earlier, weaker models. As the model improved, these instructions became less helpful. Sometimes they became the problem. Instead of solving the user’s task, the model spent reasoning effort reconciling overlapping or conflicting instructions from system prompts, skills, and the user.

The scaffolding had become friction. The design philosophy is changing.

• Replace rigid rules with principles.
• Replace exhaustive upfront instructions with progressive disclosure.
• Replace long prose with references to actual artifacts—source code, tests, APIs, schemas, and rubrics.
• Replace duplicated prompt text with reusable files.
• Replace manually maintained memory documents with automatic memory.

In other words, move information out of prompts and into the environment.

This feels like a broader trend than just prompt engineering. As models become more capable, the interface between humans and models becomes thinner. We stop explaining how to think and instead provide the information needed to think.

The codebase becomes the prompt. The tests become the specification. @file becomes more valuable than another 500-word instruction block.

This also changes how I think about the MCP server I’ve been building. Many components exist to compensate for weaknesses in previous models—routing logic, prompt templates, explicit workflows, carefully curated guidance. They were the right engineering trade-offs at the time.

But they may not remain the right trade-offs.

If frontier models continue improving, the harness doesn’t disappear overnight, but its role changes. Instead of acting as a second brain, it becomes infrastructure: fetching context, exposing tools, and getting out of the way.

Perhaps that’s the direction all AI systems are heading.