Agent 'skills' and DevEx gap
Recent tutorial videos frame agents as collections of discrete 'skills'—tool calling, retrieval, planning and memory—rather than a single monolith, and they show builders iterating on structure to improve outputs. The media analysis notes a developer‑experience gap: public tutorials teach setup step‑by‑step, which suggests enterprise platforms can capture value by packaging golden paths, templates and baked‑in observability. ( )
AI agents are being taught as bundles of small capabilities, not one all-purpose brain. Anthropic, GitHub and OpenAI now document agents as systems that combine tools, memory, handoffs and traces. (platform.claude.com, docs.github.com, developers.openai.com) Anthropic’s Agent Skills docs say a skill is packaged domain knowledge that Claude can load when relevant, and the public Agent Skills specification defines a skill as a folder with a required `SKILL.md` plus optional scripts, references and assets. GitHub’s Copilot docs use nearly the same framing, calling agent skills reusable capabilities that can include instructions, scripts, examples and resources. (platform.claude.com, agentskills.io, docs.github.com) The same pattern shows up in agent frameworks. OpenAI’s Agents Software Development Kit says agents can use tools, hand off work to specialized agents and keep a full trace, while Anthropic’s Agent Software Development Kit documents subagents as separate instances for focused subtasks and parallel work. (developers.openai.com, openai.github.io, platform.claude.com) Memory is also being broken out as its own layer. LangGraph’s memory docs describe short-term memory as thread-scoped state saved with checkpoints, and Microsoft’s Foundry docs separate long-term memory from turn-by-turn runtime state. (docs.langchain.com, learn.microsoft.com) That decomposition is showing up in how builders teach one another. Anthropic published a public `anthropics/skills` repository that had more than 116,000 GitHub stars as of April 15, 2026, and its guide to building skills focuses on planning, structure, testing and distribution rather than on a single master prompt. (github.com, resources.anthropic.com) The gap is less about raw model access than about assembly. Visual Studio Code, Claude and OpenAI all expose primitives for skills, subagents or tracing, but the public material still teaches developers to wire pieces together step by step. (code.visualstudio.com, platform.claude.com, openai.github.io) That leaves room for platforms that package the wiring. OpenAI’s tracing docs say traces are enabled by default in its Agents Software Development Kit, LangSmith sells observability around execution traces, and LiveKit markets agent observability with transcripts, session traces, metrics, logs and recordings. (openai.github.io, langchain.com, livekit.com) The commercial play is becoming clearer in documentation itself. GitHub says skills work across Visual Studio Code, Copilot Command Line Interface and its coding agent, Anthropic ships pre-built skills for document work, and Microsoft documents tracing hooks for LangChain, LangGraph and OpenAI Agents Software Development Kit inside Foundry. (docs.github.com, platform.claude.com, learn.microsoft.com) The result is a market that looks more like cloud tooling than chatbot wrappers. As agent builders keep swapping `SKILL.md` files, subagents, memory stores and tracing backends, the winners may be the vendors that make those parts feel preassembled instead of hand-built. (agentskills.io, platform.claude.com, openai.github.io)
