Autoresearch

In the optimization guide, we discussed a manual workflow for optimizing data loading pipelines: instrument the pipeline, run experiments, analyze metrics, form hypotheses, apply changes, and repeat.

This process is systematic and methodical, which makes it a good candidate for automation. Autoresearch is an engine that automates this entire optimization loop. It uses a coding agent (Claude or Codex) to analyze pipeline metrics, identify bottlenecks, propose code changes, and iteratively improve performance with minimal human intervention.

• How It Works
  • Overview
  • Getting Started
  • How It Works
    • Instrumentation
    • Seed Experiments
    • Iterative Optimization
  • Monitoring
    • Generating a Report
  • Stopping and Resuming
    • Modifying the Queue
  • Workdir Structure
  • Example
• Example: Video Classification
  • Setup
  • Baseline
  • Results
  • Key Discoveries
    • What worked
    • What did not work
  • Progress
  • Hypothesis Tree
  • Remaining Headspace
  • Experiment Statistics
• Architecture
  • Design Principles
  • Async Work Engine
  • Workflow
  • Platform Capabilities
  • Stateless Agent Invocations
  • Hypothesis Tree