Anthropic's message threads beta lets agent loops stop re-uploading the full conversation on every turn, cutting median time to first token about 35% on 500K-1M token requests. The catch: your token bill does not move, threads expire after five idle minutes, and zero-data-retention or HIPAA workload
Docker's open Sandbox Kit Spec v3 puts an AI agent's network rules, credentials, and tools inside an ordinary OCI image, and the diff becomes the security review. Here is why that beats both containers and IAM roles pasted into CI.
One team cut their measured production LLM hallucination rate from 15% to 1.5% without touching the model. The fix lived in the platform layer: a schema that doubles as a lie detector, a three-class retry loop with hard caps, and prompts you can roll back without a redeploy.
For decades, API design was straightforward: humans click buttons, call endpoints, and read responses. The consumer had a graphical interface, could tolerate latency, and would retry manually if something broke.