sharp
The article only gives an RSS snippet, so the evidentiary base is thin: the title says AI agent scaffolding is commoditizing, but the body discloses no cases, costs, benchmarks, dates, or concrete split between Claude Code, Codex, Cursor, and OpenCode. I agree with the direction. I do not buy the lighter framing that this is mainly about prompt tricks being absorbed by models.
Models and product surfaces have absorbed a lot of low-level scaffolding. The 2023 LangChain-style agent demo was mostly prompt templates, tool descriptions, ReAct loops, memory wrappers, JSON parsing, and retry logic. Claude Code and Codex CLI changed that developer expectation. You no longer hand-write a dozen system prompts to make a model plan, act, observe, and patch. Tool calls, file edits, diff generation, test execution, and context compaction are increasingly product defaults. Cursor got there from the IDE side: Tab, Composer, and Agent mode turned code-context gathering into ambient infrastructure.
But I have a real reservation about saying generic runtimes are simply being eaten. Demo runtimes are being eaten. Production runtimes are not gone. Claude Code can work inside a repo, edit files, and run commands. Codex can attach to code tasks. Cursor can operate inside the editor with rich context. That makes “I am building a general agent framework” a much worse pitch than it was in early 2023. Once this enters a company workflow, the unresolved parts are permissions, audit logs, rollback, data boundaries, queues, cost ceilings, and failure attribution. The snippet gives no reproducible case, so it does not prove these tools have replaced the runtime layer inside real organizations.
The outside pattern is already visible. LangChain spent a lot of energy moving from “agent framework” toward LangSmith because framework APIs became hard to monetize. Observability, evaluation, traces, and replay sit closer to enterprise budget. LlamaIndex also moved away from the simple “put documents into a vector database” story and toward data connectors and workflows. OpenAI’s shift from Assistants API toward Responses API also pulled tool use, files, and state management deeper into the platform. This is not a fresh Chinese-market observation. The framework companies have already admitted it through roadmap changes.
So I half-buy the line that human work becomes boundary judgment. Low-level prompting should be outsourced. Asking engineers to keep hand-crafting retry prompts, XML tags, and tool-schema nudges now feels like hand-rolling an ORM. Domain judgment, though, does not preserve itself. Teams will confuse “the model completed the task” with “the boundary was correct.” That is where the risk sits. A code agent can produce a patch that passes tests without understanding compatibility, migration risk, customer-specific deployments, or rollout policy. An agent can turn a Jira issue into a PR without knowing which abstraction should remain untouched.
I care more about the brakes than the scaffolding. The snippet does not discuss permission models or where human review enters the loop. Claude Code’s strength is its proximity to the developer loop, and that is also the danger. Once command execution and file mutation are trusted by default, the blast radius exceeds a bad chatbot answer. Cursor has the same issue: a wrong batch edit inside an IDE is harder to clean up than a wrong paragraph. If OpenCode is leaning on open source, its advantage will be control and inspectability. Its problem is that it lacks the closed-loop product data that Claude Code or Codex can collect at scale.
The title has the right instinct, but the material is not hard enough. No pricing means we cannot tell whether commoditization is a real price collapse. No benchmark means we cannot tell whether “absorbing runtime” refers to task success or just developer vibes. No task conditions means we cannot tell whether this holds in toy repos, solo projects, or million-line monoliths. Practitioners should not walk away with “scaffolding has no value.” The sharper read is: thin scaffolding has no value; execution layers with audit, permissions, evals, and recovery still do. Prompt craft is depreciating. Boundary design is appreciating. The companies that survive here will not sell an agent loop. They will make failure observable, bounded, and reviewable.