Building Around the Builders
The arrival of autonomous coding agents is creating a need for new infrastructural support layers.
Anthropic launched Routines for Claude Code last week, and Coder launched Coder Agents in beta on the same day. The two products come from unrelated companies, but they address the same transition. Coding agents that until recently required a developer’s active supervision are beginning to run unattended, on schedules and in response to repository events. That shift demands two infrastructure layers that did not exist a year ago, one to automate execution and another to govern it.
Coding round the clock
Running Claude Code on a recurring task previously meant leaving a machine online or managing cron jobs. Routines eliminate that requirement by packaging a prompt, repositories, and connectors into a saved configuration that executes on Anthropic’s managed cloud. A Routine can fire on a recurring schedule, in response to an HTTP API call, or when a GitHub webhook detects a repository event such as a new pull request. Each run creates a full cloud session with access to the repository’s skills, connectors, and shell. CI auto-fix, announced at Code w/ Claude, applies this mechanism to failing builds. When a pull request breaks CI, Claude Code opens a session against the repository and files a fix before the developer who opened the PR has reason to check.
Anthropic reported at the event that API volume on the Claude platform has grown 17x year-over-year. Mercado Libre, which employs 23,000 engineers, said that it is targeting 90% autonomous coding by Q3 2026. Early Routine patterns include scanning merged pull requests weekly for documentation drift, triaging new issues nightly, and running bespoke code review checklists on every incoming PR. Routines ship with daily execution caps that range from five runs per day on Pro plans to twenty-five on Team and Enterprise accounts. Those caps, combined with an extra-usage credit system for overages, suggest that Anthropic is treating unattended execution as a primary operating mode rather than an experiment.
The walls go up
Unattended execution creates a governance problem that attended sessions masked. When a developer runs a coding agent interactively, the developer’s judgment provides a continuous check on what the agent accesses and produces. Routines, scheduled agents, and CI-triggered workflows remove that check. Coder’s research found that 70% of companies deploy coding agents on infrastructure never designed to support them, with most relying on cloud-hosted orchestration that sends source code, prompts, and model interactions to third-party services. In practice, developers install agent extensions in personal IDEs, share API keys through environment variables, and route prompts through services whose data handling policies the organization may not have evaluated. Unattended operation, where agents run overnight or on every pull request without human oversight, makes that arrangement harder to defend.
Coder Agents runs the entire agent system, including the control plane, orchestration engine, and execution environment, on infrastructure owned by the customer. The platform connects to any model provider or to self-hosted models without intermediary routing and supports fully air-gapped deployments for organizations that cannot allow network traffic to cross their perimeter. Coder raised $90 million in a Series C led by KKR in April, and CEO Rob Whiteley said that enterprises have been “forced to choose between adopting AI agents and maintaining control” over their data. The company’s broader argument is that differentiation in AI-assisted development is migrating from the agent to the infrastructure beneath it as model capabilities converge and the agents themselves become interchangeable.
Whatever happened to “learn to code”?
The two products shipped on the same day, from opposite ends of the stack, without coordination. Until recently, a developer at a terminal provided both the execution oversight and the governance. Removing that developer from the loop means that both functions need dedicated infrastructure. The two layers are complementary, because automation without governance creates unaccountable processes, and governance without automation leaves organizations managing agents at a scale that manual oversight cannot sustain. The coding agent was the product a year ago. The infrastructure to run it unattended and the infrastructure to constrain it have since become products in their own right.