Agents are only as good as the codebase they work in. Start with fundamentals that benefit humans too: comprehensive READMEs and architecture docs, tests that describe expected behavior, clear module boundaries, strong typing, and fast builds with informative error messages. In a large legacy codebase, do this one module at a time and scope the agent to the module you’ve prepared, rather than pointing it at the whole repo.
The ROI on all of this work is far higher now, because while a new human team member onboards only once, an agent onboards every single session. Agents need explicit context that humans carry in their heads: steering files that define conventions and coding standards, skills that load task-specific procedures only when they’re needed, scripts that automate environment setup, and command line tools and MCP servers to retrieve context and take actions.
Get the agent to document its own work for future sessions. Agents should document design decisions and make liberal code comments that serve as persistent memory. The next agent session inherits the reasoning behind the code, not just the code itself.
Agents are only as good as the codebase they work in. Start with fundamentals that benefit humans too: comprehensive READMEs and architecture docs, tests that describe expected behavior, clear module boundaries, strong typing, and fast builds with informative error messages. In a large legacy codebase, do this one module at a time and scope the agent to the module you’ve prepared, rather than pointing it at the whole repo.
The ROI on all of this work is far higher now, because while a new human team member onboards only once, an agent onboards every single session. Agents need explicit context that humans carry in their heads: steering files that define conventions and coding standards, skills that load task-specific procedures only when they’re needed, scripts that automate environment setup, and command line tools and MCP servers to retrieve context and take actions.
Get the agent to document its own work for future sessions. Agents should document design decisions and make liberal code comments that serve as persistent memory. The next agent session inherits the reasoning behind the code, not just the code itself.