Agent Ops: new stacks, proofs, and runtime safeguards

Trunk Tools’ stack cut document review from 60 days to 10 by ditching general-purpose models. Trunk replaces general LLMs with a perception‑semantics‑agents stack and agent‑ready knowledge graphs, cutting review time from 60 to 10 days and making agent workflows repeatable and debuggable — a concrete pattern for Legible Landscapes and The Graph (Principles 06,11,09).

Meta’s AI chief says new Muse Spark update will sharpen coding, agentic AI. Muse Spark “Watermelon” targets stronger coding and agentic capabilities, shifting the tradeoffs between hosted model power and in‑org orchestration; outcome engineers should reassess where judgment, orchestration, and CI gates live as these models move into enterprise stacks (Principle 09).

MAS partners industry to develop safeguards for AI agents in finance. MAS and industry publish SAFR, a runtime framework that enforces policy‑bound execution, real‑time validation, and auditability for financial agents — a template for building enforceable runtime gates and audit trails into agent deployments (Principles 10,14).

Leanstral 1.5: Proof Abundance for All. Mistral open‑sources a 6B proof‑oriented model that sets SOTA on proof benchmarks and surfaces real bugs, giving teams an accessible verification model to generate machine‑checkable artifacts and tighten outcome validation loops (Principles 07,16).

Claude, please stop trying to memorize random crap. The piece shows session transcripts often degrade agent performance and argues for storing distilled artifacts and human‑reviewed documentation instead of raw chat logs — a pragmatic rule for shipping legible artifacts, preventing brittle session memory, and keeping audits tractable (Principles 08,13).