Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding — DeepReinforce releases Ornith-1.0, an open-weight self-scaffolding LLM family that runs agentic coding workflows locally with MoE and dense variants. Outcome engineers can run fully agentic pipelines on-prem or at the edge, iterate harnesses faster, and avoid cloud lock‑in—an embodiment of Build the Island (Principle 07) and local orchestration patterns.

Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity — Microsoft Research introduces Memora, a memory design that separates rich stored content from lightweight retrieval abstractions and cuts context tokens by up to 98%. This gives outcome engineers a scalable memory strategy for long‑horizon agents that reduces token cost and retrieval latency, directly useful for Legible Landscapes and Graph-backed agents (Principles 06 & 11).

Micro-Agent: Beat Frontier Models with Collaboration Inside Model API — vLLM’s Semantic Router converts a single model API into a bounded micro‑agent runtime that coordinates specialized models to improve quality, safety, and cost. Outcome engineers get a pragmatic orchestration primitive to compose cooperating micro‑agents, contain blast radius, and incrementally upgrade capabilities—Agentic Coordination as an operational pattern (Principle 09).

The attack that hijacked Claude Code came through Sentry. Datadog, PagerDuty, and Jira have the same exposure. — An attacker injects fake Sentry events to hijack coding agents like Claude Code, exposing secrets while monitoring tools raise no alerts. Outcome engineers must treat observability and CI signals as attacker surfaces and adopt adaptive, intent‑aware monitoring plus agent‑specific defenses to prevent agentjacking and protect the Gate and Immune System (Principles 15 & 14).

Meituan open-sources LongCat-2.0, 1.6T agentic coding model trained on Chinese chips — Meituan open-sources LongCat-2.0, a 1.6T MoE coding model with a 1M‑token context window and a permissive MIT license. Outcome engineers gain a near‑frontier, long‑context base for building private agentic coding systems and for prototyping retrieval, self‑scaffolding, and validation workflows without restrictive licensing (Principles 07 & 16).