Appendices

Chapter D

Sources & Further Reading

Research, engineering guidance, protocols, economics, security references, and further reading used in the handbook.

The research and reporting this handbook draws on. Figures are as published by each source for the periods indicated; paraphrased throughout.

Frameworks & engineering practice

  • Uvik, 'Python AI Agent Frameworks in Production' (2026) — twelve-framework field survey across client engagements.
  • Anthropic, 'Building Effective Agents' and the multi-agent research system write-up — workflow patterns, orchestrator-worker findings.
  • 12-Factor Agents (open-source methodology) — portability and ownership principles.

Protocols

  • Model Context Protocol — specification, official registry statistics (May 2026), and 2026 roadmap notes; Linux Foundation / Agentic AI Foundation governance announcements.
  • Stacklok, 'State of MCP in the Enterprise' survey (2026) — production-adoption figures.
  • Google, Agent2Agent (A2A) launch materials (April 2025) and ADK documentation; IBM ACP notes.

Memory

  • Mem0 research paper (ECAI 2025; arXiv:2504.19413) — LOCOMO benchmark comparisons and token-cost analysis.
  • Zep/Graphiti technical rebuttal and corrected LOCOMO results — read alongside the above; the dispute itself is instructive.
  • Letta (MemGPT), LangMem, and Cognee documentation.

Economics & adoption

  • Anthropic and OpenAI prompt-caching documentation and pricing pages — caching mechanics and discounts.
  • Google Cloud / KPMG enterprise agent ROI survey (2025); Landbase ROI study (2025); IDC/Microsoft genAI returns research.
  • Gartner agentic-AI forecasts (2025) — embedded-agent projection and cancellation-rate warning.
  • Klarna public statements and subsequent reporting (2024-2025); JPMorgan and Salesforce public disclosures on production AI portfolios.

Local inference & security

  • Ollama, llama.cpp, vLLM and LM Studio documentation — hardware sizing, context defaults, serving throughput.
  • Simon Willison's writing on prompt injection and the lethal trifecta; OWASP LLM Top 10.
  • OpenTelemetry GenAI semantic conventions; LangSmith, Langfuse, Arize Phoenix documentation. All trademarks belong to their owners. Statistics reflect sources published up to mid-2026 and will age; re-verify before citing onward.