AI agents in software delivery

When this playbook applies

Adopt this play when engineering leadership wants AI assistants to amplify delivery without sacrificing safety or trust. Typical signals include bloated review queues, repetitive operational toil, and pressure to "use AI" without a plan. The playbook keeps experimentation focused on measurable outcomes tied to DORA metrics.

Target outcomes

• Time-to-merge drops for well-scoped changes while change failure rate stays flat.
• Engineers spend less time on status reporting, summarization, and rote boilerplate.
• The organization has a clear governance model for prompt management, data handling, and auditability.

Core plays

1. Define policy and guardrails. Publish acceptable-use guidelines covering data privacy, IP, attribution, and human-in-the-loop expectations. Register bot accounts with least privilege and log every action.
2. Pick high-leverage workflows. Start with repeatable tasks such as PR summarization, release note generation, or test scaffolding. Document definition of done and what success looks like for each workflow.
3. Stand up enabling platform. Choose the runtime (CLI, chatops, internal agent platform), connect to Git, issue trackers, and documentation sources, and store prompts/version history in Git alongside code.
4. Instrument performance. Benchmark cycle time, review latency, escaped defects, and developer sentiment before rollout. During pilots, collect qualitative feedback weekly and compare metrics to baseline.
5. Iterate and scale responsibly. Tune prompts, expand coverage, and add new workflows only when existing ones show sustained wins. Share learnings through internal demos and update the developer handbook.

Operating cadence

• Weekly pilot sync to review metrics, safety incidents, and developer feedback.
• Monthly steering group with security, legal, and engineering leadership to evaluate expansion requests.
• Quarterly audit of prompts, logs, and access to verify compliance with policy.

Signals you are succeeding

• Cycle time and review latency improve for agent-assisted work without raising rework or production incident counts.
• Engineers voluntarily opt into new agent workflows and submit ideas backed by measurable outcomes.
• Documentation, prompts, and runbooks stay current because they live in Git and follow the same review practices as code.

Supporting resources

• AI agents checklist for rollouts and governance.
• FAQ addressing sponsorship, risk, and measurement questions.
• Manual reference: manual/03-ai-agents/index.md.