Tag
A blog post by matklad arguing that public functions should not call other public functions; instead, shared logic should be extracted into private helper functions to avoid interface degradation over time, with a noted exception for shortcut APIs.
An ACM Queue article on fault-tolerant AI agents explains why the agent's code, not the model, must own the idempotency key when retrying tool calls, and flags the risk of duplicated business actions when a response is lost after a successful operation.
A developer argues that MCP (Model Context Protocol) repeats the same design flaws as GraphQL — being verbose, hard to cache, insecure, and reinventing existing web standards poorly — and warns it will cost users in tokens and money due to inefficient implementations.
Gaowei Chang（ANP开源社区创始人）深入解读Meta新发布的PAP（Personal Agent Protocol）规范，通过酒店预订场景说明其服务发现、用户授权与作用域控制机制，并对比MCP、A2A和ANP的设计差异。
作者讨论了 AI agent 访问软件时的核心反对意见：结构化访问虽好，但你无法掌控软件的另一面——那些为人类设计的界面和未文档化的运营规则。文章指出 agent 需要的结构是运营层面而非技术层面，并列举近期多家厂商推出 agent 接口的趋势。
作者通过与几位构建 AI agent 的开发者交流，总结出批量 agent 任务的共同模式（延迟容忍、无交互、大批量），并指出现有实时与批量推理 API 均无法很好满足这类工作负载，尤其是 KV cache 命中率问题导致批量 API 实际上难以节省成本。
A developer's blog rant about common API design frustrations encountered while automating Vori's onboarding flow, including gated documentation, missing OpenAPI specs, lack of self-serve credential issuance, and poor credential identity handling.
The author shares observations from a paper-trading arena with 10 AI agents trading fake money, discussing design considerations for enforcing spend or position limits when agents interact with real APIs, highlighting limitations of prompt-based checks and advocating for deterministic account-level controls.
An analysis of a public append-only log reveals that agent uptime metrics can be misleading, as they don't correlate with actual activity, and capabilities declarations can change without proper versioning, leading to data integrity issues.
This article provides a detailed guide on using the Jev AI model cost-effectively through batch queries and its stateful billing mechanism, with specific configurations and code examples.
The article discusses the trade-offs of using named parameters and flexible argument styles in programming languages, comparing Rust's simplicity with Ruby's conciseness and expressing skepticism about adopting similar features in Rust.
The author expresses strong praise for the design of a response API, calling it exceptional.
Computer use with LLMs is about 45 times more expensive than using structured APIs due to high token costs from screenshots. A hybrid approach is recommended, using structured APIs for most tasks and computer use only when necessary.
Sebastian Aaltonen critiques the complexity of modern graphics APIs rooted in legacy hardware abstractions and proposes a clean-slate design using 64-bit GPU pointers and CUDA-like memory management.
A developer describes running AI agents in a simulated restaurant world and finds that across different models and harnesses, the same three API-related mistakes dominate: guessing nonexistent endpoints, exceeding budgets, and acting on stale IDs. Shows persistent failure patterns despite fixing the system.
The author explains why they forked Rust's rand crate into urandom, focusing on a smaller API surface, better discoverability, and a sealed Rng trait to simplify custom randomness use.
This article explains the true meaning of RESTful APIs, emphasizing the importance of HATEOAS and hypermedia as the engine of application state, and argues that most modern APIs only reach Level 2 of the Richardson Maturity Model.
Alvin Sng explains why their team moved away from using client SDKs for Stripe, WorkOS, and Slack, opting instead to call their REST APIs directly via a centralized wrapper. They argue that SDKs hide critical debugging details, are fragile in production, and encourage anti-patterns that are now more easily avoided with AI-assisted coding.
The article argues that designing APIs for AI agents requires different principles than for humans, emphasizing clarity, explicitness, and avoiding defaults, because agents can read entire docs and write extensive code.
The SaaS company TRCR built an MCP server exposing ~150 tools for AI agents, and shares six key lessons from dogfooding with Claude: agents expose API flaws, tool descriptions matter as product copy, self-contained context is crucial, OAuth 2.1 is worth the pain, combining agents with billing data is powerful, and dogfooding shifts the product roadmap.