Google proposes modular prompt transpilation to solve AI agent scaling challenges

When development teams first deploy AI agents, a single instruction file works adequately. But scaling to production reveals a fundamental problem: that single file becomes a dumping ground for safety rules, domain-specific logic, formatting constraints, and escalation protocols—turning it into an unmaintainable mess that breaks reliability.

Google's latest post on building scalable AI agents identifies the core issue: system prompts face the same maintainability crisis that plagued early monolithic software, but the industry hasn't yet adopted the engineering discipline needed to manage them at scale.

The three failure modes of scaling prompts#

Without proper structure, teams encounter predictable problems. Blast radius obscurity makes it impossible to reason about the consequences of changes—a single added sentence can silently break workflows across the entire system. Copy-paste drift emerges as different teams duplicate shared logic like PII handling or safety policies, creating inconsistencies and maintenance nightmares. Deferred runtime errors occur when teams resort to ad-hoc string formatting, only to discover broken variables or missing imports when rare workflows trigger in production.

These issues aren't solved by simple templating; production systems need deterministic builds, static validation, and CI/CD integration.

Treating prompts as build artifacts#

Rather than authoring one sprawling prompt file, teams should write modular skill files—each containing a specific behavior. A high-level agent prompt then composes these fragments through includes, macros, and variable injection.

Google provides an example of an SRE agent template:

# agents/sre_agent.prompt.md (prompt template file)
{% include "shared/safety.prompt.md" %}
{% include "shared/tool_usage.prompt.md" %}
You are an SRE triage agent operating in the {{ environment }} environment.
{% if allow_remediation %}
You may recommend remediation steps, but destructive actions require human approval.
{% else %}
You may inspect, summarize, and explain the issue, but do not recommend remediation actions.
{% endif %}