Coasty’s post on computer use agents vs API-only automation in legacy enterprises makes the case against the maintenance treadmill: RPA bots built on CSS classes, IDs and xpaths stop matching when a banking portal or procurement system ships a new release, and a developer rebuilds the bot. API-only automation, it says, “handles the clean cases, but legacy ERP, mainframe, and Citrix environments rarely expose stable APIs that cover the full workflow.” Its answer is an agent that reads the screen, clicks and types, and its sharpest line is this one: “A standard operating procedure written in plain English is already almost a prompt.”
Two numbers carry the argument, and neither is Coasty’s own measurement. “Up to 40 percent” of an RPA team’s time going to maintenance is attributed to unnamed analysts. The “50 percent or more” reduction in manual work is a threshold the post invents for deciding when to expand a pilot, not a result. What survives without them is the split it recommends, and it is more honest than the headline: keep RPA for “stable, high-volume, backend tasks,” use computer use agents for changing UIs and exception-heavy work, and retire brittle bots one process at a time. That is an argument for both, not against APIs.
The catalog adds the irony. Coasty on apis.io is an API company. Its 8 API pages are the control plane for exactly the pilot the post describes: the Coasty Machines API and Coasty Sessions API provision the virtual desktops, the Coasty Workflows API is where a repeatable procedure lives, the Coasty Runs API and Coasty Schedules API execute and repeat it, and the Coasty Predict API is the screen-reading step itself. The agentic access profile counts 72 operations, 42 acting and 5 human-in-the-loop. The agent that replaces the missing legacy API is driven by a well-described modern one.
The measurement backs that up. Coasty scores 62.5 on the Kin Score, strong, with access clarity at 84.2, operational transparency at 68.4 and developer ergonomics at 67.3. Agent Readiness is 48.7, agent-native, with idempotency, error semantics, OpenAPI examples, MCP server and agent skills all lit. The gap is contract governance at 4.5, and on the agent side reversibility and dry-run mode are unlit. For a product whose pitch is acting on legacy systems that cannot be rolled back, a documented way to rehearse a run before it clicks anything is the missing piece.