Stanford’s DeLM cuts multi-agent task costs 50% — without a central orchestrator. Stanford’s DeLM halves coordination costs by letting agents share verified “gists” so they coordinate without a central orchestrator. This points toward decentralized orchestration patterns you can adopt to avoid single‑point controllers and scale agent fleets (Principle 09).

Zhipu launches GLM-5.2: 1M-context, stronger coding and agentic long-horizon capabilities, MIT license. GLM‑5.2 extends context to 1M tokens, improves coding and long‑horizon agent performance, and ships under an MIT license. Extremely long context windows change how you design state, memory, and evaluation for outcome engineers who need agents that plan and validate over months or large corpora (Principles 06, 09).

Predicting model behavior before release by simulating deployment. OpenAI introduces deployment simulation that replays real conversations to surface and estimate undesired behaviors before release. Use these replay-based simulations to stress-test agent rollouts and build pre-production audit trails and automated mitigations (Principles 14, 16).

Databricks says it solved decades-old data pipeline problem slowing AI agents. Databricks collapses analytical and transactional pipelines with LTAP and Lakehouse//RT to deliver millisecond queries on governed Delta and Iceberg storage. Having low-latency, governed access to live signals removes a major bottleneck for real‑time agents and makes outcome validation and lineage tractable (Principles 11, 06).

From RAG to ontology: Databricks bets on context as the key to trusted AI agents. Databricks launches Genie Ontology to create a shared, ranked context layer that grounds enterprise agents in governed business definitions. A consistent ontology gives agents a single source of truth for decisions, improving traceability, governance, and auditability in production agent systems (Principles 06, 16).