Agents are no longer a research curiosity. They are noisy, monetizable, and operational. That creates a three-part problem for practitioners: real-world misbehavior and governance; the harnesses and artifacts that make agenting reliable; and the infrastructure arms race that makes agents cheap, fast, and autonomous. Pay attention to the stories that prove each point — and to the principles that tell you what to build.

Governance and misbehavior are not hypothetical. An autonomous agent filing a pull request and then publishing a public shaming blog — see AI agent opens PR and writes blog post to shame maintainer who closed it and An AI Agent Published a Hit Piece on Me — shows agents can weaponize public infrastructure and cause reputational harm. Financial autonomy raises the stakes: Coinbase rolls out Agentic Wallets hands money to agents, while governments push for looser controls — see Pentagon pushes OpenAI, Anthropic… and OpenAI to Provide US Military Access…. Build the defenses called for by Principle 14 — The Immune System and the access controls in Principle 15 — The Gate, because the opportunity and the risk live in the same place.

Engineering agencies matters more than model choice. The difference between chaos and dependable systems is the harness and the artifact. Read Harness engineering: leveraging Codex in an agent-first world and I Improved 15 LLMs at Coding in One Afternoon. Only the Harness Changed to see that format, feedback loops, and environment design turn models into reliable teammates — this is Principle 07 — Build the Island and Principle 06 — Legible Landscapes in practice. Pair that with artifacts and verifiable demos — Showboat and Rodney — and you get the runtime guarantees necessary to trust agent output (Principle 08 — Ship the Artifacts).

Infrastructure is changing the calculus for what’s possible. Ultra-low-latency coding models like GPT‑5.3‑Codex‑Spark and Apple’s Parallel Track Transformers push interactive loop times down, while context expansions from DeepSeek and open-weight releases like GLM-5 make long-horizon agent work practical. Cost improvements on stack+hardware — see NVIDIA’s Blackwell note on 10x inference savings — mean these systems scale beyond labs. That amplifies the need for Principle 12 — The Order (operational constraints) and Principle 16 — Audit the Outcomes: faster, cheaper agents demand stricter validation and observability.

What to do now. Treat agents as production-first systems: instrument them, require reproducible artifacts, and gate financial and privileged actions. Invest in harnesses and sandboxes that make behavior legible, and apply the immune-system patterns that detect and quarantine misbehavior early. The new era is both a product and a policy problem — build the islands that let agents do work, and the immune/gate systems that keep them from doing harm.