Tightening Agent Autonomy — Security, Open Models, and Real-Time Ops

Enterprises winning with AI agents limit how much the agents can do alone. The piece reports that successful enterprises constrain agent autonomy with rules and human checkpoints to reduce risk and integration costs. Outcome engineers should bake these human-in-the-loop controls and policy layers into agent workflows (Principles 10, 15).

One Pull to Wipe Them All. A malicious PR tried to weaponize a VS Code extension into an agent-driven wiper, forcing emergency patches and human checkpoints. This exposes supply-chain and coding-agent attack surfaces — you must add CI gating, signed extensions, runtime guards and rapid incident rollbacks (Principles 15, 14).

NanoGPT Speedrun Frontier. Agents executed 153 autonomous NanoGPT optimizer speedruns across 18 frontier models, closing parts of the human record gap. Treat this as a proof that agents can orchestrate large-scale experiments: design observable experiment lanes, replayable artifacts, and orchestration roles (Principles 09, 16).

GLM-5.3 (open-weight) beat Anthropic/OpenAI models — for 1/5 the cost. The open-weight GLM-5.3 outperforms top proprietary models on 28 tasks while running at roughly one-fifth the cost. That shifts procurement and ops calculations—evaluate open models for cost, latency, and auditability and adapt your validation and immune-system controls accordingly (Principles 14, 16).

Why real-time AI at scale is so hard. The article shows stale features, lock contention, and decaying vector indexes are the dominant failure modes for real-time AI. Outcome engineers must prioritize feature freshness, index health monitoring, and tail-latency engineering to keep agentic systems reliable in production (Principles 06, 14).