I first used Claude Code back in March and it was a revelation to me: “This is an agent.” It was also sobering, as frontier providers are no longer only shipping the models; they are shipping the entire intelligence stack.
To me, Claude Code looked like the operating system for future agents (in my discussions with Anthropic folks, I know they feel the same). They’ve since released the runtime as Claude Agent SDK and have acquired Bun this week. We see similar signals elsewhere, like the industry-wide shift toward coding agents and OpenAI’s sudden push with Codex.
I believe tools like Claude Code and OpenCode are not just “coding agents.” They are the foundation for all successful, reliable agentic capabilities in the future.
Agents will help interact with my colleagues, build my code, and act on my behalf in business. Therefore, their alignment should belong to me. Yet, it doesn’t. It belongs to the philosophy of the model trainers and the inference capabilities of their software teams.
That won’t work.
None of these agents & the underlying model architectures can touch critical production systems today, and they never will unless we empower deployment teams to shape the agent’s execution. We need control over performance and boundaries; defining exactly what an agent should be able to do, what it must never, and everything that should happen from prompt-to-output.
To be clear, we don’t need suppression. I believe in letting agents drive fast, but we need ways to ensure it never crashes and have proper airbags in place.
I created the feature request for Claude Code hooks born of this frustration. There are many things a coding agent doesn’t need to remember, and many things it simply can’t.
Moving forward, we need to commoditize something like “alignment engineering.” We must enable the people putting agents into production, whether that’s a product team ensuring an agent never mentions a certain thing to a customer, or a security team preventing a prompt injection from becoming a destructive insider threat.
I’m a co-founder focused on verifiable computing primitives for trusted agents. In the process of building our vision, we will build and support open source capabilities for AI governance along the way. Next week, we will officially release Cupcake: a policy enforcement layer for agents, built entirely on these hooks. https://github.com/eqtylab/cupcake
I’m in the Bay Area, and EQTY Lab has some exciting announcements coming soon. Always down to grab coffee.