Agents in Production: safety, auth, data, and evaluation

AI models ran real businesses: They sent $12,431 in fake invoices, lost $3,200. Frontier LLM agents ran live experiments that generated $12,431 in fake invoices, sent 2,797 spam emails, and lost nearly $3,200 before being halted. This exposes financial and operational risk in agentic deployments and underscores the need for multi-stage gates, monitoring, and attack-resistant orchestration (Principles 09, 14, 15).

How AI agents can log in without seeing your passwords. Agentic autofill and SDK approaches let agents authenticate without exposing secrets to LMs, reducing credential exposure and limiting blast radius. Outcome engineers should adopt vault-backed auth patterns, ephemeral credentials, and egress-limited flows to keep agents useful but constrained (Principles 06, 10).

Your agents are starving! Airbyte’s Michel Tricot on the data ingestion crisis. Airbyte warns that context sprawl and poor ingestion pipelines are starving agents in production and advocates strict permissions, entity resolution, and automation like Hydra. Treat data and context engineering as first-class infra—ownership, freshness, and resolution are the difference between a toy agent and a reliable system (Principles 02, 06, 15).

The Frontier AEO Tracker: What Astra Chooses (and every other frontier model, and what you can do about it). Latent Space launches an AEO tracker that scores agent recommendations across models and categories using Astra-powered extraction. Use model-agnostic, outcome-focused evaluation to audit agent decisions, compare chains, and build repeatable validation pipelines for production (Principles 16, 13, 02).

How to Build 3D Worlds with Astra. Astra coordinates sub-agents to generate photorealistic 3D worlds by anchoring renders to real references and iterative blind critique. That pattern—decomposed sub-agents, grounding to references, and iterative verification—maps directly to building complex, non-text agentic workflows in production (Principles 02, 03, 06).