Skip to content

README.md

Latest commit

 

History

History
35 lines (30 loc) · 1.73 KB

README.md

File metadata and controls

35 lines (30 loc) · 1.73 KB
Raw

DriftlessAF agentic AI infrastructure

This directory contains tools and agents for AI use with DriftlessAF.

They can be used independently but are best used together with the DriftlessAF workqueue found in /workqueue and the reconciler bots found in /reconcilers.

The following components are available:

  • AI executors: Production-ready executors for Google Gemini and Anthropic Claude models in agents/executor/.
  • System One client: Typed-question client for TypeSafe AI's Jev model or self-hosted Hopper (yes/no probabilities, choices, rubric scores) in agents/executor/systemone/. Use WithHopper or a Hopper route to select Hopper, which requires an explicit endpoint and no API key and sends one question per request.
  • Batch prediction client: Submits Claude and Gemini requests as Vertex AI batch prediction jobs and collects their results, for work that can wait hours at about half the online price, in agents/executor/batch/. A caller-chosen submission id makes resubmission return the existing job.
  • Evaluation framework: Testing and monitoring agent quality with comprehensive metrics in agents/evals/.
  • OpenTelemetry metrics: Built-in observability for AI operations in agents/metrics/.
  • Prompt building: Utilities for constructing and managing prompts in agents/promptbuilder/.
  • Tool calling: Helpers for function/tool calling with Claude and Gemini inagents/toolcall/.
  • Result parsing: Structured output extraction from model responses in agents/result/.

For providers that support the Responses API, prefer responsesexecutor. Use executor/openai/chatcompletionexecutor for providers that require Chat Completions.