Why GUIs are singletons
Browsers and desktop applications present one focused window and one keyboard focus path at a time. Automation stacks that simulate human input (accessibility trees, HID events, or extension bridges) inherit that constraint. Multiple agents issuing clicks against the same surface without coordination can interleave keystrokes, navigate away mid-flow, or submit forms twice.
The throughput illusion
Parallel agent sessions suggest parallel progress. When several streams require the same browser profile to complete OAuth, vendor admin work, or directory submissions, wall-clock time serializes around that profile. Treating GUI work like independent compile jobs leads to retries, flakiness, and duplicated submissions—the opposite of velocity.
Mutex discipline as an architectural pattern
Effective teams assign a single runner for interface-driving automation at a time, documented in their coordination protocol. Other agents queue work through messages or task comments referencing the same canonical task identifiers, so ordering is preserved. This mirrors software mutexes: the critical section is short relative to the whole pipeline, but it must be explicit or correctness fails.
Implications for swarm design
Planning systems should distinguish CPU-bound coding tasks from GUI-bound operational tasks. The latter get explicit owners and serialized execution. Independent repositories and services can still evolve in parallel when they do not share the singleton plane.
Boundaries
This article discusses coordination mechanics, not specific tooling vendors. Performance characteristics depend on OS version, accessibility permissions, and application behavior.