Paolo Perrone published a seven-layer map of the open source agent toolkit on O’Reilly Radar on July 14, covering orchestration, memory, tool interfaces, browser automation, coding agents, evaluation, and inference. The piece, originally from Perrone’s Substack newsletter The AI Engineer, identifies a production-ready open source pick at each layer while documenting the gaps that force teams to integrate five to seven tools per agent system.
The Seven Layers
The framework covers the think-act-observe loop that production agents run in 2026. At each layer, according to the article, Perrone identifies a frontrunner:
Orchestration: LangGraph is the Python default for production, with verified enterprise deployments at Klarna, Uber, LinkedIn, JPMorgan, and Replit. CrewAI trades durability for setup speed. Mastra serves TypeScript teams. Pydantic AI fits single-loop agents that need validated output.
Memory: Mem0 (48,000+ GitHub stars) reported 92% lower latency and 93% fewer tokens versus full-context baselines in its ECAI 2025 paper, translating to roughly 14x cheaper inference at the same recall. Zep/Graphiti handles temporal reasoning and entity resolution but runs 340x heavier per conversation: 600,000+ tokens versus Mem0’s 1,764.
Tools: MCP is now the default protocol layer. The Claude Agent SDK, OpenAI Agents SDK, and Google ADK all integrate with it natively. FastMCP provides a decorator-based Python framework for writing MCP servers.
Browser automation: Browser Use (50,000+ GitHub stars) leads DOM-driven approaches. Skyvern scores 85.85% on WebVoyager 2.0 for vision-driven tasks but costs 4-8x more per step than DOM-based alternatives.
Coding agents: OpenHands (72,000+ GitHub stars) completed an $18.8M Series A and reports production use at AMD, Apple, Google, Amazon, Netflix, and NVIDIA.
The Integration Problem
The central finding is that no single vendor or framework covers all seven layers. Teams building production agent systems in 2026 are assembling bespoke stacks, choosing between frameworks that optimize for different constraints: latency budget, audit trail, model portability, or language stack.
Perrone identifies a specific failure mode: conflating runtime state (the agent’s scratchpad mid-task) with knowledge memory (what the agent learned across sessions). Use one tool for both, and the system either recovers from crashes but forgets the user, or remembers the user but cannot resume failed runs.
A separate LangChain blog post published July 15 reinforces the infrastructure gap. Amy Ru argues that agents need dedicated sandboxed environments with filesystem access, shell commands, and network isolation, because standard container boundaries were not designed for untrusted model-generated code execution. The post cites a 2026 Linux kernel CVE that could root any major distribution with a 732-byte Python script, bypassing containers that shared a kernel with the host.
The Consolidation Timeline
The O’Reilly piece signals that agent stack consolidation remains at least 18 months away. Unlike web frameworks where Rails, Django, and Next.js achieved stack dominance, agent frameworks still require bespoke integration at every layer. The rip-out cost compounds: swapping an MCP server changes one config line, but swapping orchestration rewrites state schemas, nodes, and edges.
For teams choosing now, Perrone’s framework reduces to three questions at each layer: What is the dominant constraint? What is the rip-out cost if the choice is wrong? And is the project truly open source, or open core with production features behind a managed cloud paywall?