Seawater blog
Establishing the vocabulary.
Before governance is a product, it is a conversation. These posts define the terms — trust boundary, architectural airgap, earned authority, conduct engine, execution quality of service — that agent infrastructure needs to be discussed precisely.
-
2026-04-15MMX-CLI plus Routines means multimodal agents now run on a cron, with no human present and no governance layer. The concrete scenario the industry is sleepwalking into — and why consequence-tier gating and earned authority are the missing vocabulary.Live
-
2026-04-22Prompt-level guardrails are advisory. In-process middleware shares memory with the thing it's supposed to constrain. A real boundary lives in a different process — and that is a design choice, not a feature request.Soon
-
2026-04-29Trust is a demonstrated property, not a declared one. The conduct engine pattern — batting averages, regression detection, automatic de-escalation — and why this is the only model that scales to thousands of agents.Soon
-
2026-05-06If the dangerous code path exists in the binary, it can be reached. Compile-time feature exclusion — what it is, why it matters, and why "runtime policy gates" are fundamentally weaker.Soon
-
2026-05-13A walkthrough of what a governance-tested agent looks like on paper. Decision frames as evidence. Morey data, Arena scoring, OSCAL export.Soon