← back
fighting slop with slop — Vaibhav Gupta, Boundary
Takeaway
Replace exhaustive manual code reading with strong design review, architectural invariants, and continuous agent-driven testing.
Summary
- Boundary uses a small architecture.md containing durable invariants while letting engineers choose their own AI tools.
- Human attention shifts toward design documents, supported by versioning, Slack notifications, and required readers.
- Dependency visualizations and CLI checks enforce architectural boundaries in CI.
- Agents continually generate BAML programs, inspect execution traces, identify defects and inefficiencies, and test fixes or language features.
bamlcode-qualityagent-testing
Original description
You cannot tell great engineers what to do, and you increasingly cannot tell what an agent did either, so Vaibhav Gupta's answer is to fight slop with slop. At Boundary the team turns the same cheap, sloppy generation loose as a tool: agents that run constantly over the transcripts of other agents, flagging hallucinations, spotting which tool calls produced errors, and comparing which approaches produced fewer. He pairs that with hard invariants, the design docs, rules, and CLI checks that do not change for months and tell you exactly where a codebase stops converging, so the messy detection layer sits on top of something stable. The deeper move is to attack the foundational layer from first principles. Instead of trusting generated code, he leans on type systems that make whole classes of mistakes impossible: types get inferred without you writing them, a division by zero is guaranteed to be handled or the code will not build, and there are no silent unknowns left for an agent to guess at. That is the bet behind BAML, which lets you work across Python, TypeScript, or Rust with strong boundaries around each function so an agent can move fast inside walls it cannot breach. His closing challenge is to go build these sloppy tools yourself and constrain the systems underneath them, because that is what actually wins the war on slop. Speaker info: https://x.com/vaibcode / vaigup - / @boundaryml Timestamps: 0:00 - Fighting slop with slop 0:27 - Code reviews and invariants 2:08 - Building rules that don't change 3:00 - Design docs with notifications 4:03 - CLI tools that catch where things break 5:08 - Agents reading agent transcripts 5:57 - Detecting hallucinations and errors 9:24 - Attacking the foundational layer 11:03 - Execution traces from first principles 14:26 - Type safe tools across platforms 15:42 - Making whole error classes impossible 17:57 - BAML across languages 20:37 - Go build the sloppy tools