Skip to main content

​
TL;DR

  • Research Agents: Automated prompt optimization in sandboxed environments
  • Language-Agnostic Prompt Optimization: OpenAPI contracts enable Task Apps in any language
  • Multi-interface access: Web Dashboard, Python SDK, and CLI for Research Agents
  • Complete polyglot examples in Rust, Go, TypeScript, and Zig

​
Research Agent: Automated Prompt Optimization

Research Agents run in sandboxed environments to automatically analyze, modify, and optimize your prompts and workflows.

​
Key Features

  • MIPRO Integration: Agents apply MIPRO (Model-based Instruction Proposal & Refinement Optimization) to systematically improve prompt performance
  • Code-Aware Optimization: Agents understand your existing codebase, DSPy pipelines, and evaluation logic
  • Iterative Improvement: Runs multiple optimization iterations, tracking metrics and selecting best-performing candidates

​
Multi-Interface Access

Web Dashboard: Create and monitor Research Agent jobs from usesynth.ai/research
  • Visual job kickoff with GitHub repository integration
  • Real-time agent timeline showing optimization phases
  • Artifact viewer for optimized prompts, diffs, and reports
Python SDK: Full typed SDK for programmatic job management
  • ResearchAgentJob, ResearchConfig, MIPROConfig dataclasses
  • HuggingFace dataset integration via DatasetSource
  • TOML configuration support for reproducible experiments
CLI: Command-line interface for CI/CD pipelines
  • synth agent run --config research.toml --poll
  • synth agent status <job_id>

​
Research Workflow Support

  • Create New Agents: Start from scratch with a task description and dataset
  • Improve Existing Pipelines: Point to a GitHub repo with existing DSPy code
  • Benchmark Scenarios: Pre-configured for common tasks (Banking77, Iris, HeartDisease)

​
Language-Agnostic Prompt Optimization via OpenAPI Contracts

Prompt optimization now works with Task Apps written in any programming language, not just Python.

​
OpenAPI Contract

  • Complete OpenAPI 3.1 specification (synth_ai/contracts/task_app.yaml) defines the HTTP interface
  • CLI access via synth contracts show task-app or synth contracts path task-app
  • Code generation using openapi-generator for type-safe bindings

​
Complete Polyglot Examples

  • Rust: Fast, type-safe implementation using Axum framework with Tokio runtime
  • Go: Zero external dependencies, single static binary (~8-12MB)
  • TypeScript: Works with Node.js, Deno, Bun, and Cloudflare Workers
  • Zig: Minimal binaries (~1-5MB), zero runtime dependencies

​
OpenAPI ↔ Pydantic Validation

  • Automated schema validation via GitHub Actions CI
  • Prevents schema drift before integration issues
  • 19 unit tests ensure contract integrity

​
Documentation

  • SDK Reference: Complete guide at /sdk/research-agent
  • CLI Reference: New /cli/agent documentation
  • Product Guide: New /product/research-agent explaining web dashboard workflow
  • Polyglot Task Apps Guide: Complete guide at /prompt-optimization/polyglot-task-apps
  • Contracts CLI Documentation: New CLI reference at /cli/contracts

​
Technical Details

  • Sandboxed Execution: Each Research Agent job runs in an isolated Daytona workspace
  • Spend Controls: Configurable max_agent_spend_usd and max_synth_spend_usd limits
  • Model Selection: Support for GPT-5.1 Codex, Synth models, and custom configurations
  • Contract Module: New synth_ai/contracts/ module providing programmatic access to OpenAPI specifications

​
Use Cases

  • Intent Classification: Optimize prompts for multi-class classification (Banking77, support tickets)
  • Data Extraction: Improve extraction accuracy from unstructured text
  • Code Generation: Enhance code completion and generation prompts
  • RAG Pipelines: Optimize retrieval and generation prompts together
  • Existing Codebases: Integrate Synth prompt optimization without rewriting in Python
  • Performance-Critical Tasks: Use compiled languages (Rust, Go, Zig) for CPU-intensive evaluation logic