Over the last few weeks, both OpenAI and Anthropic have pushed hard on enterprise agents and long-running workflows. To me, this reinforces the shift from prompting to testing. If your agent can run longer, touch more systems, and take more autonomous actions, weak tests stop being a quality issue and become an operational risk.

OpenAI’s February releases point directly at production agents: Introducing OpenAI Frontier on February 5, and then the Stateful Runtime Environment in Amazon Bedrock plus the broader OpenAI and Amazon strategic partnership on February 27. The pattern is clear: more state, more orchestration, more enterprise deployment.

Anthropic’s February updates echo the same direction. Claude Sonnet 4.6 emphasizes gains in coding, computer use, and agent planning, while Anthropic also published Responsible Scaling Policy 3.0 and announced it acquired Vercept to push computer-use capabilities further.

When both labs are shipping toward longer-horizon execution, the bottleneck moves to feedback quality. A clever agent loop without hard checks is just a faster way to be wrong.

For me, a strong test spine for agentic workflows must be integral to the CI/CD process.

This is the key mindset: treat agents less like chat interfaces and more like distributed systems with non-deterministic components. You still need prompt quality, but your real leverage comes from measurable, repeatable, adversarial testing.

The monthly takeaway is simple. Model capability is accelerating, and reliability won’t come from capability alone. It will come from teams that build dense feedback loops and test harnesses strong enough to keep autonomous workflows inside safe, useful boundaries.

Citations: OpenAI Frontier (Feb 5, 2026); OpenAI Stateful Runtime in Amazon Bedrock (Feb 27, 2026); OpenAI and Amazon partnership (Feb 27, 2026); Anthropic Claude Sonnet 4.6 (Feb 17, 2026); Anthropic Responsible Scaling Policy 3.0 (Feb 24, 2026); Anthropic acquires Vercept (Feb 25, 2026)

Keep reading

More essays on engineering leverage and AI systems.

The Best Agents Become Invisible The best tools eventually disappear. Not because they become unimportant, but because they become ordinary. Personal agents will follow the same path. From GTC to OpenClaw, Personal Agents Get Real This month, GTC made personal agents feel less speculative. The headline was still infrastructure, inference, and the broader AI stack. But when NVIDIA put Build-a-Claw at GTC Park and described it as a "proactive, always-on AI assistant" reachable through your preferred messaging app, the more interesting signal was distribution. AI is moving from chat tabs into persistent agents you can message from anywhere. Autonomy Is a Feedback Loop I used to talk about the while (true) agent loop as a design pattern. After reading Anthropic's engineering write-up, I now treat it as a proven operating model. They used an autonomous loop to help build a real C compiler, and the result was not a toy demo. It compiled substantial code, including Linux, across multiple architectures.

Subscribe for new essays

Get updates when new writing goes live.