# Agentic search avoids stale indexes and simplifies deployment Markdown copy of https://realalpha.app/s/agentic-search-avoids-stale-indexes-and-simplifies/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy. - Speaker: [Boris Cherny](https://realalpha.app/builders/boris-cherny/) · Creator of Claude Code, Anthropic · Created Claude Code at Anthropic and leads its team; previously spent about five years at Meta. - Host: [Dan Shipper](https://realalpha.app/people/dan-shipper/) · Co-founder & CEO, Every · Sold his first startup, Firefly, to Pegasystems; writes Every's Chain of Thought. - Show: AI & I - Episode: [The Secrets of Claude Code From the Engineers Who Built It](https://realalpha.app/v/IDSAMqip6ms/), Oct 29, 2025 - Moment: 40:57 to 45:07 of the episode (4 min), playing on https://realalpha.app/s/agentic-search-avoids-stale-indexes-and-simplifies/; the episode from there: https://realalpha.app/v/IDSAMqip6ms/?t=2457 ## Quote > “They're just really tricky to maintain because you have to continuously **re-index the code** and they might get out of date” > — [Boris Cherny](https://realalpha.app/builders/boris-cherny/) at [41:12](https://realalpha.app/v/IDSAMqip6ms/?t=2472) · AI & I ## Takeaways 1. Embedding indexes need upkeep as code and local changes shift. 2. Agentic search reduced the deployment and security burden. 3. Teams can expose their own semantic search through an MCP tool. ## About this segment Anthropic moved from embeddings to agentic search because indexes drift with code changes and widen enterprise security exposure. They say agentic search reaches similar accuracy with simpler deployment. From “The Secrets of Claude Code From the Engineers Who Built It”, Oct 29, 2025. [Watch the full episode](https://realalpha.app/v/IDSAMqip6ms/) ## Transcript From the transcript AI & I publishes with the episode. [41:01](https://realalpha.app/v/IDSAMqip6ms/?t=2461) Oh, here's a good one. Why did you choose agentic rag over vector search in your architecture and are like vector embeddings still relevant? So actually, initially we did use vector embeddings. They're just really tricky to maintain because you have to continuously re-index the code and they might get out of date and you have local changes, so those need to make it in. And then as we thought about what does it feel like for an external enterprise to adopt it, we realized that this exposes a lot more surface area than like security risk. [41:37](https://realalpha.app/v/IDSAMqip6ms/?t=2497) We also found that actually, Cloud Code is really good, and Cloud Models are really good at agentic search. So you can get to the same accuracy level with agentic search, and it's just a much cleaner deployment story. That's really interesting. If you do want to bring semantic search to Cloud Code, you can do so via an MCP tool. So if you want to manage your own index and expose an MCP tool that lets Cloud Code call that, that would work. [42:04](https://realalpha.app/v/IDSAMqip6ms/?t=2524) What do you think are the top MCPs to use with Cloud Code? Ooh, Puppeteer and Playwright are pretty high up there. Definitely, yeah. Sentry has a really good one. Asana has a really good one. Do you think that there are any power user tips that you see people inside of Anthropic or other people who are big power, inside of organizations that are big cloud code power users that people don't know about, but they should? [42:35](https://realalpha.app/v/IDSAMqip6ms/?t=2555) One thing that CloudCo doesn't naturally like to do, but that I personally find very useful, is CloudCo doesn't naturally like to ask questions. But, you know, if you're brainstorming with a thought partner, a collaborator, usually you do ask questions back and forth to each other. And so this is one of the things that... I like to do, especially in plan mode, I'll just tell QuadCode, like, hey, we're just brainstorming this thing. [43:00](https://realalpha.app/v/IDSAMqip6ms/?t=2580) Please ask me questions. If there's anything you're unsure about, I want you to ask questions, and I'll do it. And I think that actually helps you arrive at a better answer. There's also so many tips that we can share. I think there's a few really common mistakes I see people make. One is, like you said, not using plan mode enough. This is just super important. And I think this is people that are kind of new to agentic coding. [43:24](https://realalpha.app/v/IDSAMqip6ms/?t=2604) They kind of assume this thing can do anything. And it can't. It's not that good today. And it's going to get better. But today, it can one-shot some tests. It can't one-shot most things. Yeah. And so you kind of have to understand the limits and you have to understand like where you get in the loop. And so like something like plan mode, it can like two, three X success rates pretty easily if you like land on the plan first. [43:44](https://realalpha.app/v/IDSAMqip6ms/?t=2624) Other stuff that I've seen power users do really well is companies that have really big deployments of quad code. And now, you know, luckily there's a lot of these companies so we can kind of learn from them. having settings JSON that you check into the code base is really important because you can use this to pre-allow certain commands so you don't get permission prompted every time and also to block certain commands let's say you don't want web fetch or whatever and this way as an engineer I don't get prompted and I can check this in and share it with the whole team so everyone gets to use it I get around that by just using dangerous they skip permissions Yeah, we kind of have this there, but we don't recommend it. [44:23](https://realalpha.app/v/IDSAMqip6ms/?t=2663) It's a model. It can do weird stuff. I think another kind of cool use case that we've seen is people using stophooks for interesting stuff. So stophook runs whenever the turn is complete. So this is in some tool calls, back and forth, whatever, and it's done. And it returns control back to the user. Then we run the stophook. And so you can define a stophook that's like, if the tests don't pass, return the text, keep going. [44:49](https://realalpha.app/v/IDSAMqip6ms/?t=2689) And essentially, it's like you can just make the model keep going until the thing is done. And this is just insane when you combine it with the SDK and this kind of programmatic usage. This is a stochastic thing. It's a non-deterministic thing. But with scaffolding, you can get these deterministic outcomes.