Still wrestling with flaky AI pilots and surprise production incidents? This brief shows what changed, who moved when, and what you should do next. Late‐Dec 2024–early‐2025 saw LLMs shift from one‐off calls to orchestracted agent workflows in production—Salesforce (12/23, 12/27), HubSpot (12/22, 12/28), DoorDash (12/28) and Shopify (12/30) run agents over CRMs, ticketing and observability with human checkpoints. Platform teams centralized AI (Humanitec 12/22; CNCF 12/23; Backstage 12/27–12/28). Security and policy tightened: CISA urged memory‐safe languages (12/22) and SBOM work advanced (Linux Foundation/OpenSSF 1/02/25). Apple (12/23) and Qualcomm (12/30) pushed on‐device models. Observability vendors (Datadog 12/20; Arize 1/02/25) tied LLM traces to OpenTelemetry. Immediate takeaway: treat agents as platform products—standard APIs, identity, secrets, logging, and human gates before you scale.

Worried about latency, privacy, or un‐auditable AI in your products? In the last two weeks vendors shifted multimodal and compiler work from “cloud‐only” to truly on‐device: Apple’s MLX added optimized kernels and commits (2024‐12‐28 to 2025‐01‐03) and independent llama.cpp benchmarks (2024‐12‐30) show a 7B model at ~20–30 tokens/s on M1/M2 at 4‐bit; Qualcomm’s Snapdragon 8 Gen 4 cites up to 45 TOPS (2024‐12‐17) and MediaTek’s Dimensity 9400 >60 TOPS (2024‐11‐18). At the same time GitHub (docs 2024‐12‐19; blog 2025‐01‐02) and JetBrains (2024‐12‐17, 2025‐01‐02) push plan–execute–verify agents with audit trails, while LangSmith (2024‐12‐22) and Arize Phoenix (commits through 2024‐12‐27) make LLM traces and evals first‐class. Practical takeaway: target hybrid architectures—local summarization/intent on-device, cloud for heavy retrieval—and bake in tests, traces, and governance now.