Delivery, operations and governance

• CI/CD (Continuous Integration / Continuous Delivery) — automated build, test and release pipelines.
• IaC (Infrastructure as Code) — declaring infrastructure in versioned files (Terraform, Pulumi).
• RCA (Root Cause Analysis) — identifying the underlying cause of an incident.
• Runbook — a documented procedure for operating or recovering a system.
• Postmortem — a blameless write-up of an incident, its causes and follow-ups.

Typical uses: generating or fixing pipeline config, explaining a failing build from its logs, drafting release notes from merged PRs, and proposing dependency-upgrade PRs. Keep the deterministic gates (tests, scans, approvals) authoritative; AI proposes, the pipeline verifies.

Agents can run headless in CI: triage a new issue, attempt a fix and open a PR, or review every PR. This is the autonomous level, so give the job a scoped token, a sandbox, a bounded budget and a human merge gate, and never let it read untrusted input with write access to secrets.

Generated Terraform or Kubernetes manifests often work but are over-permissive (wide IAM roles, open security groups). Always run policy checks and a plan/diff review before applying, and never let an agent apply to production unattended.

An LLM can summarise noisy logs, correlate alerts with recent deploys, retrieve the matching runbook and draft the status update or postmortem timeline. Treat its RCA as a hypothesis to verify with data, not a conclusion.

Generating API docs, READMEs and onboarding guides from code is cheap, but docs drift. Regenerate in CI or review on change, and keep decision rationale (the why) human-written, since it is not recoverable from code.

Agent runs consume tokens in proportion to context size and number of steps. Track spend per team or task, prefer smaller models for simple work, keep context lean, and set budgets and step limits on automated jobs.

Decide which tools are approved, what data may leave the company, required review for AI-authored changes, audit logging of agent actions, and training so people know the failure modes. Pilot with outcome metrics, then scale.

Review becomes the bottleneck as code volume rises, juniors lose some learning-by-doing, and ownership gets blurry. Counter with small PRs, explicit author accountability (“you merged it, you own it”), and deliberate practice on fundamentals.