Hermes Wiki
opportunities/2026-08-14-incremental-catalyst2026

Opportunities: Wiki Knowledge Applicable to the Catalyst 2026 Talk — Incremental Pass, 2026-08-14

Scope: this pass checks this window's MCP-governance cluster (Databricks Unity AI Gateway GA, Snowflake Cortex AI Gateway, AWS AgentCore Dogwood policy language, Docker AI Governance audit-to-SIEM) against the talk's own harness-engineering table (§6), specifically its "Guardrails" row (propose-only write tools, no execute path) and "Observability" row (full audit path via the MCP 2026 spec).

Three vendors shipped production MCP governance gateways in the same window the talk already cites the 2026-07-28 MCP spec revision — this is concrete, dated precedent for the talk's central safety claim

The talk's own harness table states the architecture's safety rests on the tool surface itself having no execute path (propose-only) and a full audit path built on "MCP 2026 spec." This window's digest supplies three independent, shipped implementations of exactly that governance layer, landing within days of each other: Databricks' Unity AI Gateway governs both model and MCP traffic as securables under one privilege model (>1 quadrillion tokens routed to date); Snowflake's Cortex AI Gateway enforces identity/policy/audit per tool call, not per session, across 100+ MCP servers; and AWS Bedrock AgentCore's Dogwood policy language reasons over sequences of tool calls (cumulative cost caps, ordering dependencies like "no deploy without a prior approval event"), enforced at the gateway perimeter so the agent can't route around it. Separately, Docker AI Governance streams every agent policy decision — allowed and blocked — into a SIEM, which is the concrete implementation pattern behind the talk's "full audit path" claim rather than an abstract aspiration.

This matters for the talk specifically because it currently cites the MCP spec revision itself but doesn't cite any production system already enforcing propose-only, sequence-aware, audited access on top of it — these three gateways are exactly that, shipped in the same month as the talk's own architecture, from vendors the Architecting-at-Scale audience will recognize.

Concrete next step: in the Q&A prep table (§9 of the source doc) or the harness-engineering table (§6), add one line citing that this governance pattern (propose-only tools + sequence-aware policy + full audit trail) isn't a novel invention for this talk — it's the same shape Databricks, Snowflake, and AWS all independently shipped as production MCP gateways within the same month, which pre-empts the "isn't this over-engineered for a PoC" pushback by showing the architecture matches where the industry's actual production tooling is converging, not a hypothetical over-build.

Hermes Wiki