DX now needs machine legibility
As AI tools move from writing code to running workflows, designers of internal libraries are being urged to make side effects explicit, error states inspectable and APIs predictable so both humans and agents can use them correctly. That guidance appears alongside calls to use agents as first customers to tighten API specs and documentation. (9to5mac.com) (x.com)
Software teams are starting to treat internal libraries like infrastructure for robots, not just coworkers, as coding agents move from autocomplete to scheduled execution. (code.claude.com) Anthropic put that shift into product form on April 14, 2026, when it introduced Claude Code “routines,” a research preview that lets users save a prompt, repositories, and connectors, then run them automatically. Those routines can fire on a schedule, through an Hypertext Transfer Protocol post request, or from GitHub events. (code.claude.com) (9to5mac.com) Anthropic said the runs happen on its own cloud infrastructure, so they continue when a laptop is closed. The company limited the feature to Claude Code on the web for Pro, Max, Team, and Enterprise plans, with daily caps of 5, 15, and 25 runs depending on plan. (code.claude.com) (9to5mac.com) That changes what a “developer user” looks like inside a company. A human can guess what a helper function probably does from naming and convention; an agent that is opening pull requests or responding to alerts needs the side effects, triggers, and failure conditions spelled out in a form it can inspect. (openai.com) (nordicapis.com) The same pressure shows up in the tooling around these systems. Anthropic’s managed-agents documentation describes a hosted runtime where Claude can read files, run commands, browse the web, and execute code, while OpenAI’s agents documentation describes applications that plan, call tools, and keep state across multi-step work. (platform.claude.com) (developers.openai.com) Once software is being called by an agent instead of a person, vague behavior becomes an operational bug. A library that silently writes to disk, retries in the background, or collapses several failure modes into one generic error can still be usable for a human engineer and still be hard for an agent to recover from correctly. (nordicapis.com) (github.com) Anthropic has already been changing Claude Code to support longer-running autonomy. In an October 20, 2025 engineering post, the company said sandboxing reduced permission prompts by 84% in internal use by constraining filesystem and network access, which let Claude act with fewer approvals. (anthropic.com) That is pushing a new standard for developer experience inside companies: explicit side effects, structured outputs, stable schemas, and error messages that say what failed and what to do next. Microsoft, Salesforce, and other vendors are now publishing separate guidance for agent-facing extensions and “agent experience” alongside traditional user-facing design. (learn.microsoft.com) (salesforce.com) One result is that agents are becoming a first customer for internal platforms. If a routine, agent, or tool call cannot reliably use an application programming interface without custom glue, the problem is often not the model alone but the interface contract, the documentation, or the hidden state in the system. (openai.com) (nordicapis.com) As more code moves from “assist me” to “run this while I am away,” the old test for a good library is getting tighter. It now has to make sense to a person reading the source and to a machine deciding what to do next. (code.claude.com) (platform.claude.com)
