newsfilter
← Interviews
AI Engineer13 min

Every Correction You Give Your Agent Is Data - Brian Douglas, Paper Compute

Brian Douglas

Conference Presentation, Talk
  • Core Thesis: Engineers should explicitly label "pushback" events (instances where humans correct AI agents) to capture institutional knowledge that AI lacks, rather than verbally abusing agents without recording the outcome.
  • Junior Engineer Impact: The proliferation of agents to handle routine "reps" (repetitive tasks) risks depriving junior engineers of the necessary engagement required to level up to senior roles.
  • Open Source Tooling: The speaker maintains an Apache 2.0 open-source project called Tapes (tapes.dev) designed to collect and store agent session data and traces.
  • Runtime Environment: A custom runtime tool named Stereos is used to create reproducible sandboxes and environments that generate specific signals during agent interactions.
  • Parallel Scaling: The speaker's team utilizes a "sweeper agent" capable of running approximately 10 parallel agents, previously scaling to 200 concurrent Daytona sandboxes for defined tasks like documentation, testing, and code cleanup.
  • Personal Constraint: Despite daily usage of Rust in production at their company (Paper Compute), the speaker admits to not knowing how to write Rust professionally, having only done it for side projects a decade ago.
  • Pushback as Signal: Sessions marked with "pushback" indicate a high-confidence signal where a human (e.g., "Matt") possesses domain context the agent lacks, such as specific API structural decisions or knowledge of existing service duplications.
  • The Review Gap: Junior engineers relying on agents often lack the context to push back early, resulting in Pull Requests (PRs) that receive zero human review commentary but fail when checked by senior engineers who know the system's nuances.
  • Twin Reviewer Concept: To solve the review gap, the speaker created a "Twin Reviewer" that uses captured traces to simulate the review style of specific senior engineers on command line interfaces.
  • Data Isolation: The Twin Reviewer relies exclusively on internal traces and post-training data; it intentionally excludes public GitHub PR review data to maintain internal privacy.
  • Post-Training Strategy: The speaker performed specialized fine-tuning (adapters) on a Qwen 4B model using captured "golden sessions" (marked pushbacks) and regressions to teach the model organizational coding standards.
  • Avoiding Reinvention: The training specifically targets common agent failures, such as agents reinventing existing Go libraries (e.g., mocking libraries or Kubernetes concepts) due to a lack of context.
  • Direct Preference Optimization (DPO) Constraints: DPO, described as "CliffsNotes" compared to "margin notes" of fine-tuning, requires large models (70B+) and significant compute resources (e.g., NVIDIA GPUs on Brev), making it inaccessible for small local setups compared to specialized fine-tuning on 4B-6B models.
  • Recommendation Path:
    • Start by building skills and using Tapes or similar capture tools.
    • Use the captured data to create evaluations.
    • Apply specialized fine-tuning on smaller models (e.g., Qwen 4B) before attempting DPO on larger models.
  • Community References: The speaker references Rick Rubin's The Creative Act regarding "vibe coding" and an article by Addy Osmani (Anthropic) regarding the loss of junior engineering growth opportunities.
  • Tool Availability: While Tapes is open source, Paper Console (the centralized trace storage) is currently wait-listed; the speaker notes TwinReviewer is not open-sourced due to sensitive internal data.