Paolo Perrone Maps 2026 Open-Source AI Agent Stack Across 7 Layers

2 articles · Updated · O'Reilly Media · Jul 14

Perrone argues the 2026 open-source agent toolkit has solved most production pain points, but in incompatible ways that force teams to choose tools by their dominant constraint rather than by benchmark rankings.
Four constraints drive most picks across the seven layers—latency budget, audit trail, model portability and language stack—because the cost of choosing wrong can range from a simple config change to a full state-schema rewrite.
At the layer level, he highlights LangGraph for audited orchestration, Mem0 for low-latency memory, MCP as the standard tool protocol, Browser Use or Stagehand for browser control, OpenHands for coding agents, Langfuse for tracing, and vLLM as the default inference engine.
The report stresses that benchmark leaders can hide steep production trade-offs: Zep’s temporal graph memory can exceed 600,000 tokens per conversation versus Mem0’s 1,764, while Skyvern’s 85.85% WebVoyager score still implies about 1 in 7 multistep tasks fail.
His broader conclusion is that agent stacks do not compose cleanly top to bottom; reliable 2026 systems are built as seven independent bets, with teams accepting integration work at the seams.