Buildkite adds pipeline secrets and repair tools for AI agents

Buildkite makes continuous integration software (tools that test and assemble code). It added secrets that only one pipeline can use. It also added new tools for the Model Context Protocol (MCP), a standard that connects AI agents to software. The tools let an agent check a pipeline, find the cause of a build stop, and repair a webhook.

How it works

• The validate_pipeline tool checks the pipeline file against the schema before a build starts.
• The compare_builds tool compares a build with an error to the last good build.
• The create_pipeline_webhook tool adds a missing webhook to a pipeline.

The numbers

Buildkite gave no test results for the tools. Tests by Buildkite showed that one alternative caused more tool calls by agents, and it changed the default back. Support for all Agent v3 releases ends on 3 September 2027.

What is not known

The check does not include plugin data, connections between steps or steps that the system makes during a build. A change in version 1.23.0 breaks calls from the previous version: create_pipeline must have a create_webhook value. Scoped secrets do not control how a step uses them.

Insight: Providers give agents tasks but keep approval with persons

Buildkite gives agents small tools, and a user must be able to change the pipeline and its cluster. GitHub AI Scan cannot stop a merge. Descope adds an approval step before an agent gets access. This reads as a common move: providers give agents more work but keep last control with a person. Find where each of your suppliers puts that stop.