You install Google Antigravity, start a session in a project folder and answer the trust prompt that protects you from unfamiliar contents. You then use artifacts to review a plan before the agent acts, adding a comment that sends it back for revision. You finish by letting Antigravity create your first agent, which produces an AGENTS.md file with rules and skills. Those pieces are the foundation for every guardrail you add in later modules.

You compare four ways to reduce the risks of large language models, from prompt engineering to training from scratch, and see why the cheapest option comes first. You then apply the same idea to agent output with the verification ladder: formatting, linting, tests and continuous integration, each deterministic. You see why a language's own toolchain, its verification surface, matters, and you watch a one-shot agent build a whole project whose coverage report shows files that no test ever touched.

You turn written rules into automation by adding a Makefile, so the agent can run one target that bundles every check. You then try to enforce test-driven development through the agent's rules, see Antigravity ignore the rule at first, and learn that enforcement sometimes takes correcting and restating. You finish with PMAT, a tool that goes beyond tests and linting with a quality gate, a score and a compliance report that agents can read and act on.

You see why continuous integration is the referee for agent work: a fresh runner with only what the repository declares, which can block a merge on red. You learn the parts of GitHub Actions, including workflows, events, jobs, steps and runners, then add Dockerfile linting with hadolint to pull requests and watch it flag the latest image tag. You finish by asking an agent to create pull request workflows and seeing remote checks, such as a coverage threshold, hold its own work to account.

You examine a real repository where agents do most of the work and verification gates all of it, with thousands of commits, nightly tests, PMAT configuration and file-size limits. You then ask Antigravity to analyze the repository and summarize how it verifies agent work, including pre-flight verification, ratcheting quality gates, falsification, a multi-agent review quorum and provable contracts. You leave with a way to investigate and apply these patterns to any project.