Title: Claude code web submodules issues | Dinis Cruz

Dinis Cruz documents a practical challenge that emerges when deploying Claude Code Web—Anthropic's AI agent for code generation—against real-world projects: the tool doesn't support Git submodules, a common development practice for managing modular codebases. Rather than a simple limitation, this reveals a deeper infrastructure problem: many development tools were built for human developers, not autonomous AI agents that need to understand and work with complex project structures.

Cruz's documented solution path is instructive for those implementing AI agents: it involved exploring multiple workarounds before settling on GitHub Personal Access Tokens (PATs) as a pragmatic, if imperfect, interim solution. What makes this noteworthy is not the technical fix itself, but what it demonstrates about the current state of agentic infrastructure.

The post highlights a critical observation in the accompanying documentation: "Infrastructure Not Fit for the Agentic World." This captures a broader challenge facing organizations adopting AI agents—existing systems, processes, and tools were optimized for human workflows and may create friction when autonomous systems need to operate within them. Security, authentication, access control, and code organization practices all need reconsideration when the 'user' is an AI agent.

For business professionals evaluating AI agents, this illustrates why tool selection and infrastructure assessment matters. It's not just about the agent's capabilities, but whether your existing technical environment can support autonomous operation without requiring workarounds that introduce risk or complexity.