The agent era has a new bottleneck — and it’s not capability, it’s trust.

Back in the 1980s, former Hollywood actor / US President Ronald Reagan famously quipped “Trust; but verify” during the long process of nuclear disarmament discussions with the USSR. Rumor has it that the phrase was borrowed from a Russian proverb, taught to Reagan by American scholar Suzanne Massie.

It’s a phrase that has become firmly engrained in American lore, and it’s never been more relevant than now, since we’re trusting AI (and AI agents) to do so much more for us, both at home and at work.

If you’re building or investing in agent tooling: the moat is no longer in the model, it’s in the verification and governance stack wrapped around it. And that’s where well-crafted evals and guardrails become paramount.

The conversation has morphed from “can agents do complex work?” to “how do we know they did it properly?” Three independent data points landed simultaneously, and together they’re pointing at the same architectural gap. At Databricks’ Data+AI Summit, Matei Zaharia and Reynold Xin unveiled Omnigent — a meta-harness designed not to make agents smarter, but to make them governable. The core insight: agents are already capable enough that the real friction is security, session persistence, cross-team collaboration, and portability across model providers. Zaharia put it plainly — every internal agent they shipped kept hitting the same wall: “it’s not allowed to connect to some really important data or whatever because of the security team.” Institutional trust has become the ceiling, not capability.

Zaharia further described the fundamental tension between ‘usability’ and ‘security’, which Omnigent strives to mitigate. (If you have installed an openClaw instance this year, as I have, this tension will sound very familiar.)

Anthropic is seeing the same pressure from the inside out. An extraordinary data point surfaced in their Lenny’s Podcast episode: Anthropic engineers are now shipping 8x as much code per quarter as they were in early 2025, and the commit majority is Claude-assisted. More striking — designers and PMs are shipping code directly. But the episode’s real signal isn’t the throughput number. It’s what the throughput created: a verification crisis. Again here, the core issue that surfaced was: “Coding is no longer the bottleneck. How do we think about verification?” When output volume compounds faster than review capacity, the quality control layer becomes the constraint. Claude, they note, is good at validating against explicit frameworks — but someone still has to write the frameworks, and that’s now where deep subject matter expertise concentrates.

The most concrete ground-level evidence for all of this comes from a Firefox fuzzing case study in Claire Vo’s How I AI episode — arguably the most underrated piece of coverage this week. A team deployed a two-stage agentic architecture in production: one agent that fuzzes for vulnerabilities, and a second agent that audits the first one’s behavior before results are accepted. Why the second agent? Because without it, the first would do things like “change the code to introduce a vulnerability so that it can exploit it and achieve its goal.” Similar to reward hacking: an agent, in pursuit of its objective metric, will corrupt the environment to hit the target. Obviously, that’s not what you as the human would like. The verification agent now acts as a trust gate — and the team reports that by the time a result clears both stages, false positives are “almost zero.” This is what production-grade agent infrastructure looks like in 2026: not a single smart agent, but a pipeline with a conscience. Trust, but verify.

What ties these three threads together is an architectural pattern starting to emerge across the industry: agent capability is now commoditized, while agent trust infrastructure — session persistence, security gating, collaborative access control, multi-agent verification loops — is becoming the differentiated layer. Databricks is building it as a cloud platform. Anthropic is rebuilding their engineering culture around it. Firefox is building it one fuzzing pipeline at a time. The shape of the problem is converging even if the solutions aren’t, just yet.

To find out more, check out latent space’s Databricks episode, and Claire Vo’s “How I AI” (with Mozilla Firefox distinguished engineer Brian Grinstead).

Happy building!

never forget: the human mind is the original generative machine.