AI agents left the chat box

Agents left the chat box. Anthropic’s hardware standard connects physical devices through 3 control paths—MCP, CLI, and APIs—while Google’s prediction engine divides geospatial modeling into 3 stages and 4 anti-leakage checks. One moves microscopes and robotic arms; the other builds public-health and risk models from natural-language requests. Across 21 CDC indicators, Google reports mean R-squared of 76.8% versus 60.0% for a manual pipeline. The shared shift is from generating answers to coordinating systems.

The experiments are getting a scorecard. Sapient says its PRAXIST benchmark run earned 60 medals across 75 tasks, including 49 golds, at roughly $3,054 in model spend; its Claude Code baseline logged 55 medals, 34 golds, and $38,370. Salesforce’s Claudeforce rollout takes a narrower route with 37 sales skills and one centrally managed connection. Salesforce also reports 8.1 million annualized productivity hours from Slackbot, more than twice the prior quarter. Benchmarks measure capability; permissions decide whether capability can enter production.

Scale is no longer the only constraint. At Carnegie Mellon, an MHS setup linked 3 computers, used 96-well plates, and cut integration to about 8 hours; Anthropic says the experiments ran roughly 3 times faster. Google says PPE compresses weeks of data engineering into minutes through 3 independent stages. Salesforce plans a September 2026 open beta after a select pilot. The practical divide is becoming clearer: agents need a safe interface, measurable outputs, and a permission boundary before autonomy becomes useful.

Latest from PulseMark

Snowflake Cortex Agents Coding Agent: REST API Guide Configure Snowflake’s approval defaults, role boundaries, workspace mounts, and reusable skills before the agent gets a shell. Read the tutorial →