Luisa Lima

Secure coding practices with agents

The other half of the problem: not securing the agent, but shipping secure software with one. Posture, supply chain, and the practices that used to live in engineers' heads and now have to live in the agent's instructions.

Ongoing · 6 chapters

Securing the agent is one problem, but using the agent to ship software that holds up is another. And as software engineering increasingly means “AI-assisted engineering”, it’s the problem most teams will actually have.

The practices that make software resilient used to live in senior engineers’ heads, in our collective “boo-boos” and in wikis (that unfortunately mostly nobody read).

With agents writing the code, those practices have to move into instructions, defaults, and checks that fire on their own. Discipline you have to remember is discipline you and the agents will skip when it matters.

  1. 01. The task is the variable

    Safety in AI-assisted coding is not really a dial between fast and careful. The real variable is the task: adjust your posture to match it.

  2. 02. Agents only do what you ask, and that is a problem... and a solution

    The practices that make software hold up used to live in engineers' heads and in wikis nobody read. They now need to live in instructions that steer an agent.

  3. 03. Skills are just text files. So where's the lockfile?

    Skill distribution is a mess: no manifest, no version pinning, no lockfile. A skill is a prompt injected into a privileged agent, so it's a supply-chain problem. The boring fix already exists.

  4. 04. Dependency scanning should be a default, not a discipline

    Vulnerability scanning fails like ergonomics: fine when you remember, forgotten the one time it matters. I created autoscan-kit to push the scan into places that fire on their own, so skipping it is harder than running it.

  5. 05. Accuracy over agreement: my standing instructions for AI agents

    The custom instructions I've kept on every AI agent since Claude first supported them. Opus 4.7 is the first model that actually follows them.

  6. 06. My Agent Operating Policy

    The brief I give agents I want operating with less supervision, under AGENTS.md.