Awesome AI-Native Engineering

This roadmap helps engineering teams build their agent harness step by step, from exploring a product and writing specs to running fleets of parallel agents and improving the harness with session retros. The Agentic AI Toolbox at the end lists tools by category.

Contents

• Working with AI agents
• Spec-first work
• Agent instructions, Skills, and tools
• Testing and quality engineering
• Parallel agents and autonomous execution
• Harness and context engineering
• Security, maintenance, and policy
• Agentic AI Toolbox

Working with AI agents

1. Product exploration

Use the agent to learn how the product works from its code. Ask questions such as “How does feature X work?”, “What is component Y for?”, or “Which services handle a checkout request?”. It can also map the repo, summarize the tech stack, and draw component diagrams.

In practice: your AI agents can answer the questions about the architecture, components breakdown, and key workflows.

2. Changes that pass build and tests before review

Let the agent run its own loop: change code, build, run static analysis and tests, read the failures, fix, repeat. Put build, lint, and tests behind one command that finishes in minutes and gives the same result every run. Set an iteration limit and review the final diff for edits to tests, snapshots, and check configuration, since an agent that cannot make a check pass sometimes changes the check.

In practice: build, lint, and tests run with one command in minutes, the agent loops on them with a set limit, and the final diff has no unreviewed changes to tests, snapshots, or check configuration.

3. Independent AI review before human review

An agent tends to approve its own code. Before a human reads a diff, have a context that did not write it review the diff against the spec: a reviewer sub-agent, a second agent, or a PR bot such as CodeRabbit or Greptile. Use required CI checks and branch protection so the agent cannot merge on its own.

In practice: before a human sees it, every agent diff passes CI that the agent cannot edit and a review against the spec by a context that didn’t write it, and this review has caught at least one defect in a task the agent reported as done.

4. Continuous refactoring and cleanup