When the library isn't in the training data, point the agent at the source
The most useful thing a coding agent can do isn't recall APIs from memory — it's read the source code in front of it. A small Simon Willison tip reframes what 'knowing a library' means in April 2026.
The most useful thing a coding agent can do isn't recall APIs from memory — it's read the source code sitting in front of it. That's the quiet argument Simon Willison made this week, and it's worth dwelling on because it reframes what "knowing a library" even means for a tool in April 2026.
Read, don't recall
In a short reply on X, Willison explained his workflow: "I use plenty of libraries in my work that aren't in the training data, so I tell the coding agent to read the library source code when it needs to understand them." It sounds obvious after you hear it and mildly heretical before. The default assumption — that a coding assistant's usefulness is bounded by what its base model saw during pretraining — evaporates the moment you treat the model less as an oracle and more as a junior engineer with a filesystem.
This matters because the long tail of real codebases is where most developers actually live. Private SDKs, internal forks, newly released packages, patched versions — none of that made it into the training snapshot of any frontier model as of April 2026. Reading the source isn't a workaround; it's the primary job.
The ecosystem follows
The second signal this week is less technical but no less telling. swyx noted that Cognition's Windsurf is opening a Singapore office, with AI Engineer community organizers involved in the launch. Physical real estate and developer-relations footprint are the kind of investments you make once a category has stopped being experimental. Coding agents are infrastructure now, and the companies building them are planting flags in the Asia-Pacific developer hubs that will determine day-one adoption in the next wave.
Implications
For builders, the "my model doesn't know this library" complaint is increasingly a prompting problem, not a model-capability problem. Point the agent at the source. For buyers evaluating coding tools as of April 2026, watch for products that can actually traverse a repository and read unfamiliar code — that capability separates demoware from daily drivers. For non-technical readers, the same principle scales up: treat the AI as a collaborator that needs context, not a vending machine of answers.
The bottom line: When your agent gets stuck on an unfamiliar library, don't switch models — hand it the source.