13
Fact-check residue
Six myths that failed verification
These circulate in blog posts but were refuted 0–3. Keep them straight.
"OTel has finalized AI-agent semantic conventions."
No. Still in-progress SIG work.
"Multi-agent beats monolithic by default."
No. Conditional; minimum-complexity wins.
"There are exactly 7 canonical design patterns."
No canonical fixed list.
"There are exactly 5 canonical design patterns."
Same — no canonical list.
"Multi-agent collaboration requires A2A + MCP."
Overstated as a hard requirement.
"27,100 searches proves LangGraph is most-adopted."
No. Search ≠ adoption; AutoGen leads GitHub stars.
14
Intellectual honesty
What this research did NOT nail down
Serving economicsManaged platforms (Bedrock AgentCore, Azure AI Foundry, Vertex Agent Engine) vs. self-hosted K8s/Ray/Temporal — adoption and cost largely unverified here.
Memory in practiceWhich specific vector stores and memory architectures actually dominate production, vs. the vendor-blog landscape.
Deeper case studiesBeyond headline logos — cross-industry architecture + ROI detail is thin.
Eval / guardrail maturityWhich eval frameworks teams standardize on to close the 37-point gap.