A delivery pipeline is a deterministic workflow: build, test, scan, and deploy according to rules the team can read. A language model can assist that workflow. It can summarize a failed log, draft a pipeline change for review, or explain a diff. It does not reliably “fix” a pipeline because it produced a confident paragraph. The failure may be a flaky test, a missing secret, a bad manifest, or a downstream outage. Those have different remedies.
Separate the assistant from the deploy identity
NIST’s AI Risk Management Framework treats trustworthiness as something designed into the system, not as a property of the model name. In a pipeline, that means the assistant’s credentials are not the production deploy credentials. A suggestion is a pull request or a comment. A person, or a separate policy check, merges it. If an automated repair is allowed, its permissions should be bounded: which repositories, which files, which environments, and which commands. Timeouts and an audit record of the prompt, the tool calls, and the resulting change are part of the control, not optional telemetry. NIST AI RMF
The Model Context Protocol is relevant when the assistant needs a tool, such as reading a build log or opening an issue. MCP defines how a client discovers tools a server is willing to expose. It does not authorize those tools by itself. The pipeline identity still has to be allowed to call them. MCP architecture
What is worth automating
- Classification of failures from logs, with a link to the raw log so a person can check the summary.
- Drafts of tests or manifest changes that cannot merge without review.
- Policy checks that are deterministic: dependency scanning, signature checks, required reviewers. Do not replace those with a model judgment.
Evaluation belongs here. Keep a sample of suggestions and record whether reviewers accepted them. A tool that creates more broken changes than it prevents is a cost, even if the demo looks fluent. Kubernetes will run the resulting containers. It will not review the change, and it will not supply the model. Kubernetes overview
