The Agent SDK

Original Article ↗ Hacker News Discussion ↗

Rapidly proliferating “agent SDKs” like Cursor’s are drawing comparisons to JavaScript framework churn, with many engineers skeptical that ever-more elaborate agent workflows beat well-crafted prompts and simpler tools. Commenters weigh the trade-offs between proprietary harnesses such as Claude Code, Cursor, and OpenAI’s offerings—often tightly tuned to their own models and subsidized via subscriptions—and open-source or custom harnesses that promise flexibility but may lack deep integration and stable economics. Underneath the branding and acronyms, the core tensions are vendor lock-in, the absence of clear benchmarks or standards, and whether complex multi-agent architectures deliver real productivity gains or just productivity theater.

Ecosystem, Standards, and Vendor Lock‑In

  • Several comments want a common standard (e.g., OpenResponses) so apps aren’t tied to specific vendors (OpenAI, Anthropic, Cursor, etc.).
  • Others note we’re in a “pre‑standards” phase: lots of parallel attempts because no one knows the “right” design yet.

Agent SDKs, Harnesses, and Abstraction Levels

  • Some argue an “agent SDK” could be as simple as PROMPT.md + runtime + tools; complex frameworks like LangChain feel overengineered.
  • Counterpoint: runtimes add needed control—different tools, models, permissions, guardrails, quotas, and auditability across workflows, especially for enterprises.
  • Several people describe using Temporal, gRPC, k8s, or Google’s ADK as orchestration backbones rather than heavyweight agent frameworks.

Skepticism: Productivity Theater and Churn

  • Many see endless “agent” workflows as hype, with little empirical evidence they beat good single prompts.
  • Comparisons are drawn to JS framework churn, OOP/UML diagram mania, and endlessly tweaking dev environments (e.g., neovim configs).
  • Some criticize multi‑agent “software factories” that generate huge unreviewed PRs, warning this degrades engineering standards and increases risk.

Open‑Source vs Proprietary Harnesses

  • A sizable group sticks with proprietary harnesses (especially Claude Code and Cursor) due to:
    • Deep integration with frontier models and subsidized subscription pricing.
    • Belief that models perform better in their “native” harness.
    • Familiar UX, low switching cost pressure, and perceived lack of standout features in OSS harnesses.
  • Others advocate open harnesses (Pi, OpenCode, custom ones), emphasizing:
    • Portability across models/providers.
    • Ability to route between subscriptions and APIs.
    • No real lock‑in, since harnesses can call each other’s CLIs.
  • There is confusion and disagreement about Anthropic’s terms, billing, and whether third‑party harnesses draw from Claude subscriptions vs separate API credits; docs and behavior are described as ambiguous but recently updated.

Cursor, “Agent SDK,” and Alternatives

  • Some feel Cursor’s UX declined as it chased agentic workflows, preferring its earlier, simpler “tabs + cmd‑K” model.
  • Others like Cursor’s agentic cloud/deploy story and see “agents” as workloads to host and monetize.
  • Comparisons mention competing tools (Anthropic Agent SDK, OpenAI Agents, Vercel AI/Eve, Devin), but the thread finds nothing obviously unique in Cursor’s Agent SDK pitch.

UI/UX and Real Business Needs

  • One practitioner says the hardest part is domain‑specific chat UIs and approval flows; generic chat interfaces and MCP/tooling don’t solve this, forcing many to rebuild frontends themselves.
  • Various libraries/frameworks (TanStack AI, ChatGPT Apps, custom TUIs) are mentioned as attempts to address richer UI flows on top of agents.