GoCD vs Buildkite

Buildkite is shipping in two directions at once, and both point at automation as the consumer. On the agent side: an MCP server that now waits on builds and summarizes failures in one call, plus broad new REST and GraphQL coverage for pipeline creation, org administration, notification services, hosted agents, and artifacts. On the pipeline side: a checkout block in pipeline YAML that absorbs what previously took Sparse Checkout, Custom Checkout, and Skip Checkout plugins, Test Engine metrics for reliability and duration behind a dated version header, and job logs exportable to an OpenTelemetry collector. Agent v4 becomes stable on 1 September.

The organizing idea is headless CI — every operation that used to require the web UI becoming callable, and every failure becoming legible to something that is not a human reading a log. Promise job failure is the sharpest expression of it: the build lifecycle itself is being reshaped so remediation can start before a job finishes. The versioned API header and the v3-to-v4 stable transition suggest Buildkite is settling its interfaces deliberately rather than accreting them, which is what you do before you expect programmatic clients to depend on them.

Expect the MCP server to keep collapsing multi-call workflows into purpose-built tools the way get_build_failure_summary and wait_for_build did, and expect promise-based signalling to extend past failure to other states a long-running job can predict.