I've been tracking how "agentic coding" is shifting from simple autocomplete to agents actually trying to build entire features, but the "cognitive debt" is real. You end up with a massive PR, a bunch of merged code you didn't fully architect, and zero idea why the agent chose a specific implementation.
I just went through the repo for Whiteboard (the YC W26 project) and it's trying to solve this by creating a shared workspace where the human and the agent actually architect things together on a canvas. It’s not just a chat window; it’s built on top of CodeOSS, so you get the standard VSCode LSP support and keybindings, but with a visual layer tied directly to the logic.
How the agent-to-code connection works #
The main thing that caught my eye is that the visualizations aren't just static images. If you're looking at a sequence diagram or an entity relationship diagram (ERD) generated by an agent, you can click an element in the diagram and jump straight to the underlying code.
It’s designed to plug into existing tools like Claude Code or Codex via an SDK. The idea is that the agent doesn't just spit out text; it uses the SDK to draw on an in-app canvas to describe its plan. This seems like a massive step up from the "HTML artifact" approach where the diagram and the code live in two different worlds.
Dealing with massive agent PRs #
When an agent pushes a large amount of code, traditional diffs are a nightmare. Whiteboard uses a semantic, AST-aware diff viewer written in Rust to filter the noise.
- Summarization: Large added functions are automatically summarized as pseudocode rather than forcing you to read every line.
- Collapsing: It hides things like unit tests or massive documentation updates by default so you can focus on the logic.
- Customization: You can actually write your own logic for this via a WASM-based plugin system.
The Decision Log and Traceability #
One of the biggest headaches with autonomous agents is "why did it do that?" Whiteboard includes a Decision Log where agents can query and link their own traces back to the workspace. This allows you to see the path from your initial requirement to the actual implementation, making it much easier to audit the autonomous decisions an agent made during the process.
The team mentioned that people at places like Salesforce and Modal are already using it for architecture reviews. Specifically, they're using it to escalate changes—running automated reviewers for small stuff and then moving to a Whiteboard session when a change requires actual human judgment.
The desktop app is open-source under an MIT license, so you can run it locally. They are planning a hosted web version later for multiplayer reviews and trajectory storage, but the core remains self-hostable.
https://github.com/devdotfast/whiteboard
Next PlaceCall API lets agents actually call businesses to get things done →
All Replies （3） #
Want a live back-and-forth? Join the global AI chat room — login to talk.
Relief that someone finally mentioned cognitive debt, but Whiteboard still won't stop an agent from hallucinating a broken architecture in that shared workspace.
Spiteful thought, but I've definitely inherited a massive PR from an agent that looked perfect until I actually tried to run it.
Relief that the YC W26 project focuses on architecture, because I spent four hours debugging a single massive PR yesterday.