Harnesses & SDKs

Python and TypeScript SDK that lets an agent message people on email, Slack, Discord, Telegram and other channels through one handler.

Python toolkit for Vietnamese NLP with a multi-provider agent module that calls LLM APIs using only the standard library.

Code-first Java toolkit for building and testing AI agents, with multi-agent hierarchies, a dev UI and Maven packages.

Python interface for running and evaluating agents from several frameworks, with OpenTelemetry traces; now in soft deprecation.

SDK for building Microsoft Teams apps and AI agents in TypeScript, C# and Python, with a CLI that scaffolds new agents.

Library for LLM post-training (SFT, DPO, GRPO and agentic RL) and classic RL, with evolutionary hyperparameter optimization.

Open-source online code execution system that runs submitted code in a sandbox behind an HTTP JSON API.

Python SDK for building agents on Volcengine AgentKit, with a bundled Web UI, Studio deployment and short- and long-term memory.

Reinforcement learning framework for training reasoning LLM agents over multi-turn episodes with the StarPO algorithm.

Open-source infrastructure for continually self-improving agents that serves requests, learns from feedback and versions updates.

YAML-driven Python robot simulator for navigation, control and learning research, with collision detection and a Matplotlib renderer.

Lightweight Python agent framework for long-running tasks, with skills, sub-agents, A2A and ACP servers, and CLI, TUI and web front ends.

Rust multi-agent framework with typed tool calling, sliding-window memory, pluggable LLM backends and OpenTelemetry tracing.

Event-sourced Python graph runtime where agent behaviors react to a shared graph, with an audit trail and fork-and-diff replay.

Open-source NVIDIA library that connects existing agents to tools and data, with profiling and LangSmith tracing; deprecated, EOL end of 2026.

Cross-platform C++ DAG framework with a Python version, for running task graphs with no third-party dependencies.

Reinforcement-learning trainer of about 3,500 lines that trains models behind agents run in their own harnesses, locally or as Kubernetes jobs.

TypeScript library that lets agents call MCP and UTCP tools by running model-written code in a sandbox.

Reliability layer for tool calling with self-hosted LLMs, offered as a proxy, a workflow runner and middleware.

Multi-interface agent framework with tools, memory and MCP, serving agents through Telegram, Discord, Twitter, Farcaster and a REST API.

Open-source Python evaluation framework with ready-made and custom graders for scoring AI agent quality and producing reward signals.

Python library of 35 agentic AI architectures with one run() interface, built on LangGraph and multiple LLM providers.

Go framework for AI agents with memory, tool execution, MCP, agent skills, multi-agent patterns and evaluation.

Framework for developing, evaluating and benchmarking web agents on BrowserGym benchmarks, with parallel runs through Ray.