Standardizing Agentic Code: Building Guidelines for AI and Human Engineers

Building shared coding guidelines for AI (and people too)

Software engineers in 2026 are increasingly utilizing coding agents to generate code based on high-level designs rather than writing it manually. Vish Abrams, chief architect at Heroku, warns that without explicit prompts for principles like DRY, LLMs may produce unmaintainable code that ignores standard configuration-code separation.

Why This Matters

While human developers absorb tacit context and ‘vibe-based’ code understanding through experience, AI agents lack this intuition and operate purely on explicit instructions. As agents generate more code, the cognitive burden shifts to design and review; without deterministic guidelines, engineers face an overwhelming review process for code they did not write, potentially leading to a ‘Tower of Babel’ service-oriented architecture.

Key Insights

• Externalizing Tacit Knowledge: Greg Foster (CTO of Graphite, 2026) notes that engineers implicitly absorb codebase context, which must be converted into objective, explicit documentation for agents.
• The agents.md Flywheel: Quinn Slack (CEO of Sourcegraph, 2026) advocates for a continuous feedback loop where agent errors are used to update rules files, creating a high-signal development flywheel.
• Deterministic Enforcement: Logan Kilpatrick (Senior Product Manager at DeepMind/Google, 2026) emphasizes that established enterprise style guides provide the ‘ripe context’ needed to prevent agents from taking shots in the dark.
• Pattern-Based Guidance: Documentation must move beyond lists to include ‘gold standard’ files that serve as end-to-end templates for AI models to emulate.
• Hybrid Tooling: Traditional linters, formatters, and static analysis tools remain essential in the build pipeline to catch basic errors that agents may still botch.

Practical Applications

• Company: Heroku-style application deployment. Use case: Explicitly instructing agents to separate configuration from code to ensure maintainability. Pitfall: Failing to specify architectural constraints, leading to agents creating monolithic ‘snake games’ that are difficult to refactor.
• Company: Sourcegraph-style agentic workflows. Use case: Implementing an agents.md file checked into a standard repository to provide context for Claude skills or other LLM agents. Pitfall: Treating agent prompts as one-off tasks (chicken-typing) rather than a persistent, version-controlled standard.

References:

• https://stackoverflow.blog/2026/03/26/coding-guidelines-for-ai-agents-and-people-too/