Platforms for AI agents: CLI, YAML turn intent into execution

The design write-up is by a practitioner who started from an existing frontend development and delivery platform. The starting point is that AI agents introduce new entry points — a user asking from inside a repository, one agent generating code and another requesting a build, or a Git event triggering a flow — and those entry points do not come with the information a person would have typed into a platform screen.

The author's response was to split two roles that had been folded into one template: choosing which procedure to run, and executing that procedure according to its dependencies while recording the result. Commands alone turned out to be insufficient because a caller cannot read screen colors or a neighbor field, so the platform was reshaped around the three capabilities of querying, starting, and checking. In practice the harder work was not adding APIs but uncovering which information existing processes silently depended on, such as change-management units that may not yet exist when a Git event is the starting point.

Beyond single-agent operation, the write-up argues that a conversation and a job have different lifetimes, so discussions, tasks, runs and dependency relations need to be tracked as separate units, with diagnosis, recovery and notification treated as distinct states. The author cautions that this is an accumulation of design choices rather than a finished result, and frames the value less in how many commands an agent can call than in whether a user can reach an outcome, explain a wait, trace a failure, and resume midway.