The Open Source Agent Toolkit in 2026

A survey of the 2026 open source AI agent ecosystem, organized into seven independent architectural layers. The core thesis: most production problems have a solution, but each has been solved a dozen incompatible ways. The dominant constraint for your system—latency budget, audit trail, model portability, or language stack—determines which tool fits.

Orchestration. LangGraph is the Python production default (durable graph execution, verbose setup). CrewAI optimizes for rapid prototyping (role-based agents, no state schema) but lacks crash recovery. Pydantic AI is best for single-loop agents returning validated data. Mastra is the TypeScript-native option for Next.js apps. Vendor SDKs trade lock-in for zero-friction setup.

Memory. The context window is not memory. Mem0 provides hybrid vector+graph memory with 92% lower latency vs. full-context approaches, but lacks temporal reasoning. Zep/Graphiti tracks how entities and relationships change over time, at higher cost. Letta (formerly MemGPT) treats memory like an OS: RAM for active context, disk for archival, with the agent managing promotion. Distinguish runtime state (mid-task scratchpad) from knowledge memory (across-session learning).

Protocols. MCP is the de facto standard, supported by every major vendor SDK and framework. FastMCP is the Python server framework; mcp-agent orchestrates across multiple MCP servers.

Browsers. Browser Use (Python, LLM-driven, 50k+ stars) is best paired with cached Playwright scripts for the repeated 80%. Stagehand (TypeScript, MIT license) uses AI only for reasoning steps. Skyvern (vision-first, 85.85% on WebVoyager 2.0) handles form-filling where DOM is unreliable. Production teams wire both DOM-driven and vision-driven paths.

Coding agents. OpenHands (72k+ stars, $18.8M Series A) is the autonomous production option with event-stream architecture, scoring 53-72% on SWE-bench Verified. Aider (35k+ stars) is Git-native, terminal-only, with an Architect/Editor mode that cuts costs 30-40%. Cline (38k+ stars) is VS Code-native with Plan/Act separation. Most teams run two coding agents: one commercial, one open source.

Evals and observability. Langfuse is the open source default for tracing LLM calls, tool invocations, and costs. Arize Phoenix is OpenTelemetry-native for existing observability stacks. Inspect AI focuses on safety evals. The article's engineering lesson: wire tracing on Day 1, before the first user.

Inference. vLLM is the production serving default for open-weight models. Ollama is for local use only. llama.cpp runs on CPU and Apple Silicon without GPU, defined the GGUF format. SGLang benchmarks faster on agent workloads but has a smaller community.

Overall: the seven layers are independent decisions, not a vertical stack. Teams shipping reliable agents pick the best tool per layer and accept that integrating the seams is part of the job.