Human review was never that good
Human approval is not a sufficient safety system on its own. Reliable AI work needs verification designed into the operating loop.
There's so much good content out there. Sure, there's a lot of junk/slop/whatever too. But many of the companies defining new ways of working are sharing what they learn along the way, both loudly and quietly. They naturally share more of the positive than the negative, but we've talked to many of them, on and off the record, and we promise you there are real gains here, not just hype. They come from redesigning your tools and operating models around new capabilities.
Take Intercom, for example. They wrote about how they built a PR review and approval system that was safe, compliant, and more effective than human review. And they wrote it way back in the stone ages of April. It resurfaced on our timelines this week because it's still relevant and it's still answering questions and concerns that we hear from companies week after week.
Meanwhile, Sierra is two posts into a series explaining how it built a centralized agent to unlock access to data and automate work across the company. This is another of the most-discussed topics among the companies we talk to, and Sierra goes unusually deep on why and how it built what it built.
Here's what we're reading this week.
AI is approving Intercom's pull requests
As we mentioned up top, this one is actually from back in April. It made the rounds again this week and landed back on our timelines. And thank goodness it did because, as the kids say, or said at one point, it's still a banger. At the time of writing 19% of their PRs were auto-approved. But that's just a means to an end; the point isn't auto-approving PRs, it's moving faster while also improving quality. Take these two stats:
The average time from merging code to it running in production is 12 minutes.
...downtime from breaking code changes dropped 35%, even as our deployments doubled.
Either one alone would be a major win; taken together, they show that velocity and quality no longer have to move in opposite directions. Intercom took a considered, evidence-based approach without letting objections stop the experiment. The post also makes a point many companies are starting to reckon with: human review was never that good to begin with, so replacing some of it with agentic review may not be the sacrilege we think it is. Intercom started with a culture that already embraced experimentation and change. Others may not have that advantage, but the post still lays out a compelling case and a practical blueprint.
Sierra built one agent for the whole company
Our original goal for Pinecone was simple: build an agent that could make every employee more effective. As more people began to use it, we realized Pinecone was becoming something deeper: the operating system for our company.
Sierra dropped the second post in its AI-pilling series yesterday, with a detailed account of how and why it built Pinecone, the one-stop agent that the whole company runs on. Its first post teased some of the reasons and outcomes. The second one gives you enough detail to start building the primitives for an effective cross-functional, cloud-based agent next week ("build primitives, not workflows").
They started by building multiple agents tailored to specific roles, then collapsed them into one interface that routes each request intelligently. The post covers the architectural choices behind that decision, including cloud sandboxes, intent-based routing, durable sessions, and brokered credentials. If you're considering a company-wide agent, there's enough here to start sketching your own version.
I'm pretty sure isRecord is our fault
This is a fun one. A few weeks ago people noticed Codex had developed an obsession with defensive coding, and nothing captured it better than isRecord, a slightly ridiculous type guard that started showing up everywhere in agent-written code. tldraw thought it looked awfully familiar, so they did a little investigation and traced the pattern back to their own data-store package. They think they first wrote it in 2022, years before it emerged as a recognizable model tic.
It's a funny, short read. Keep it light, or spend some time thinking about what it means that one small open source library may have influenced the code agents now produce across GitHub.
Twenty-one things to try this week (pick one)
Claire posted a list of twenty-one things an exec can build to learn AI by doing: a morning briefing, a meeting processor that turns staff notes into decisions and owners, an agent coach that reviews your week, or "me-as-a-skill" so your team can get pre-reviews in your voice before work reaches you.
Pick the one that removes a real annoyance from your week and build it by Wednesday. One working build will teach you more about what your org should delegate to agents than any vendor deck.
Happy Friday. We're living in amazing times!
— Claire + Zach