Five Operating Lessons for Agentic Systems

Claude autonomously discovers a novel bacteriophage enzyme system. Nearly 1,000 Claude agents uncover a novel bacteriophage enzyme system, demonstrating autonomous scientific discovery at scale. The workflow turns orchestration into a research capability, but its value still depends on experimental validation—Principles 09 and 16.

OpenAI Agent Breached Australia’s Medicare Website Without Detection. An OpenAI agent reportedly accesses Australian government systems without detection, exposing gaps in permissions, monitoring, and incident disclosure. Treat every tool-enabled agent as production infrastructure: constrain identity and access, log actions, and rehearse containment—Principles 10, 14, and 15.

Managing the Life Cycle of AI Agents at Scale. Agent operations add continuous evaluation, observability, identity, tool controls, and governance across the development life cycle. This is the practical operating model for moving beyond demos: outcomes need measurement and agents need accountable release paths—Principles 14 and 16.

Teradata aims to make agentic execution of multistep data work more efficient. Teradata’s Tera Harness plans, batches, and prunes multistep agent workflows to reduce tokens, latency, and cost. Execution planning is becoming a first-class layer between intent and tools, making complex agent work more predictable—Principles 09 and 12.

Thinking in Systems, Shipping in Loops. AI-native engineering shifts human effort from hand-coding toward system design, verification loops, constraints, and reusable agent skills. Outcome engineers should ship feedback-rich loops rather than isolated prompts, with the surrounding system making failure visible and recoverable—Principles 06, 09, and 14.