AI agents.

How to let AI agents do real work in a codebase while you keep the rules: what they may touch, what must pass, and what stays your decision.

This chapter is being outlined. Every section below is a stub with a summary; the full text is not written yet.

1. Principles Agents work inside rules you set: the scope they may touch, the gates they must pass, the decisions that stay yours, and an audit trail of what happened.
2. Lanes A lane is one stream of work with its own branch, worktree and path scope, so several agents can run at once without touching each other's files.
3. Briefs and decisions How to hand an agent work: write a brief, make it plan first, and let it hold a question for a person instead of guessing.
4. Pipelines and gates Pipelines are where rules become enforceable: ratchets that only tighten, required checks, and CI recipes an agent cannot talk its way around.
5. Code review Review that agents cannot skip: an AI review job in the pipeline, the CI approval flow, soft approvals, and an audit log of every decision.
6. Standards for agents Write the house rules once, in files agents read: AGENTS.md or CLAUDE.md, standards packs and architecture rules, backed by a machine-enforced coding standard.
7. Secrets for agents and pipelines Agents and pipelines need credentials without ever holding them in code: how uBixVault issues them and how Replikate carries them into Kubernetes.
8. Automation with Kelpie uBixKelpie is meant to pick up Git issues and fix them, and to run servers that monitor code. It is at the foundation stage: this page describes intent, not shipped behaviour.
9. Cost and models Choosing which model does which job, and putting a ceiling on spend: providers, budgets and caps.
10. Recipes Worked, copyable examples that tie the sections together: CI YAML, scripts and agent briefs, tested so they do not go stale.