Code as Data

Chapter 6. Code as Data Is the Foundation for Agentic Systems

Coding agents are now part of how software gets built. More than that, they’ve triggered the most significant transformation the software lifecycle has seen in decades. The debate about the next generation of language models continues, but practically, the conversation has focused on agent orchestration—agent selection, coordination strategy, subagents, plan mode, role-scoped workflows, MCP servers, and elaborate harnesses. However, the real conversation starts further down, at the source.

Agentic AI runs on context at scale. An agent’s output is only as good as what it can see, including not only the file in front of it but also dependencies, callers, and conventions across the whole codebase. When that signal is missing, the LLM infers and guesses. In a large enterprise codebase, the consequences compound, creating bottlenecked review cycles, verbose and needlessly complex code, confidently incomplete modernizations, and ratcheting token costs.

That’s why the LST, the code model we introduced in Chapter 2, becomes a valuable layer between the agent and the source. It holds types, dependencies, and call graphs, all resolved and queryable across thousands of repositories as a single corpus. Figure 6-1 shows how the LST fits between your codebase and the agent activity. The rest of this chapter explains what that layer makes possible.

Figure 6-1. Enabling agents to reason against the code-as-data layer, not the raw ...