Ambient Agents and 6 Other Big Ideas Coming Out of AI

0:00Today on the AI Daily Brief, seven big ideas coming out of AI. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:34let's dive into these ideas, starting with Ambient Agents. As usual for the weekend, we are doing a big think episode. And this week, instead of focusing on just one idea or one concept, I was noticing that there are a bunch of really interesting themes that I've seen starting to emerge more and more that at some point will probably all get their own episodes of the AI Daily Brief if they haven't already. So what we're going to do today is go through seven big ideas coming out of AI, talk about what they mean for AI, but also for the world at large. And we're kicking off with a concept that is certainly not new, but is coming to the fore right now, which is background or ambient agents.

1:10Now, why I think this is interesting is that it's a reframe and a reimagining of the way that agents will interact with how we do work. It is a cleaner break from the co-pilot or assistant model than some of the agents that people use and deploy on their behalf today. What's more, based on some new features from Cursor over the last month, it's something that's more in production now than it ever has been. This is a concept that people have been talking about for a while now. LangChain's Harrison Chase back in August of last year, 2024, wrote, One of the UX patterns I'm most interested in is the idea of ambient agents working in the background for us.

1:44This can be super powerful but needs some UX tricks to get right. What are some considerations? Human in the loop? We mostly talk about human in the loop, but that doesn't work for ambient agents. Still, having observability and potentially control over what the agent does is important. So rather than have human in the loop, what if they were on the loop? You can observe what is happening, but after the fact. Next UX pattern is asking for help. Oftentimes, agents can't be fully autonomous. They may need input from a human on a particular point. For example, I have an email agent which often needs input from me on how to fix this bug, whether I want to schedule this meeting, etc.

2:15This means that agents need a way to reach out to humans and ask for help. What might this look like? I think something like a customer support dashboard is a reasonable UX. So this was back before the reasoning models had even launched. Then in January of this year, Langchain took it a step farther. They wrote, Today we're releasing an open source email assistant agent that we've been using internally for the past six months. This is the first agent in a new UX paradigm we're calling ambient agents. They continue, While the dominant UX for LLM applications so far is chat, we imagine this changing.

2:44Ambient agents are agents that are always on, listening to an event feed, and responding to or flagging those events as necessary. This will let us scale ourselves far more than chat. So you're seeing here that this is not just about a different type of agent, it's about a different type of interaction with agents. In June, Lance Martin from Langchain dropped a free course on GitHub about building ambient agents. Again, he writes, this is one of the most interesting agent UX patterns, allowing the agent to do work in the background and interact with the user via human in the loop for select actions and approvals.

3:14In this clip, Langchain's Harrison Chase explains why this could be so significant for professionals. There's a few reasons why we're excited about ambient agents. So the NBA agents we define as agents that are triggered by events, run in the background, but they're not completely autonomous. And so I want to break down those things. So if they're triggered by events, that means that I can scale myself as a professional a lot more because I'm no longer kind of like kicking off an agent. There's not a one-to-one kind of like interaction. It's now triggered by events. So there could be thousands, hundreds, thousands, millions, whatever, of agents running in the background at the same time.

3:49So again, the idea here is that these are agents that don't just make us more productive, but actually fundamentally change and scale the amount of work we can do. Sean Wang, aka SWIX from Latent Space and the AI Engineering Summit writes, Ambient agents are going to completely dominate the rest of 2025. He points out that one, human deep work and focus requires at least one to two hours uninterrupted. And two, by end of year, all next gen models will pass the one to two hour autonomy barrier. Basically, they will be used in completely different ways than the current 1-15 minute autonomy frontier.

4:21So what Sean is arguing here isn't just that this is a hot trend that people are watching, but that from a capabilities perspective, we are right on the cusp of agents being able to do more extended and more complex work in a way where we can let them off and do their thing. You'll see that SWIX is a mainstay of this list, and so if you are looking for new follows on Twitter slash X and you're somehow not already following SWIX, highly recommend it. VC Jessica Liao points out the obvious but important fact. The greatest ROI for AI comes when AI is working in the background, handling a workflow rather than requiring excessive prompting.

4:52And I think there's a strong chance that once we are fully in the ambient agent paradigm, it's going to make the primarily prompting paradigm feel like the Stone Age. Google DeepMind's Philip Schmidt writes, Chat agents are successful but limited by their request response human-initiated nature. The next iteration of agents will operate proactively in the background or in ambient. Ambient or background agents will not wait for direct human command, instead will operate proactively with trigger initiated by events. Now, Philip points out that one of the key updates will also be expanded long-term memory, as that's going to be required for ambient agents to accomplish their goals.

5:25Now, as I mentioned, one of the reasons that this is interesting now is that in a recent update from Cursor at the end of June, they basically launched background agents that you could initiate from your phone. We want Cursor to be the best place to write code with an AI agent. And currently, that means working with the agent side-by-side in your editor or having it start tasks in the background for you. And today, we're bringing cursor agents to the web. Just type in a task and an agent will spin up and get to work for you, making changes to your code base, answering contextual questions, and opening up PRs on your behalf.

5:55When it's done, you can carry on the agent's work directly in cursor or add follow-up instructions with additional context and even make inline edits. And if the agent's work looks complete, you can create and merge a pull request directly from the app. So what does this look like in practice? Entrepreneur and podcaster Ryan Carson writes, I've been testing the brand new Cursor mobile background agent for almost a week, and I've been shipping a lot more code. I want to do a quick review of the new Cursor background agent on mobile. So here I have it installed right here on my home screen. Let's fire it up.

6:30And this is beautiful. So from the couch, from bed, wherever you are, you can go ahead and ask her to do stuff like this. So I previously said fix a critical bug and then fired that off. And you can choose your model here if you want. So Cloud4Sonnet is my go-to and Max. You can also upload a photo if you'd like. And then you fire that off. And this is what you get once an agent has been working. So you can see it found a bug related to the middleware. So what's important is not exactly what it's doing, but the fact that this is initiated from a mobile app, all happening in the background. It's a new mode of interaction.

7:12And another important thing is that this is not just coding. With ChatGBT's recently released agent, we're starting to get this experience as well. One of the things that OpenAI did in their announcement was basically say, this is something that's meant to run in the background while you go to other things. Yes, you can watch the chain of thought and see how it's interacting with the computer and even intercept it and take over its virtual machine if it's helpful. But mostly the idea is that you tell it to do something and then it does it. Box's Aaron Levy writes, Here's the new ChatGPT agent combing through market and data strategies stored in Box to produce a PowerPoint presentation.

7:43The ability for agents to connect to data from anywhere, use a computer, and automate work in the background is going to be wild. This is obviously a paradigm we are going to be talking about a lot more, but it's definitely hit an inflection point recently where it's worth considering on its own terms and starting to think about what it's going to mean for your own work to spin up and deploy ambient agents that run in the background. Today's episode is brought to you by KPMG. In today's fiercely competitive market, unlocking AI's potential could help give you a competitive edge, foster growth, and drive new value.

8:13But here's the key. You don't need an AI strategy. You need to embed AI into your overall business strategy to truly power it up. KPMG can show you how to integrate AI and AI agents into your business strategy in a way that truly works and is built on trusted AI principles and platforms. Check out real stories from KPMG to hear how AI is driving success with its clients at www.kpmg.us slash AI. Again, that's www.kpmg.us slash AI. This episode is brought to you by Blitzy, the enterprise autonomous software development platform with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale codebases with millions of lines of code.

8:58Enterprise engineering leaders start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and precompiles code for each task. Blitzy delivers 80 % plus of the development work autonomously while providing a guide for the final 20 % of human development work required to complete the sprint. Public companies are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native SDLC into their org.

9:28Blitzy is providing a limited time, 30-day free proof of concept for qualifying enterprises. The team will provide a 5x velocity increase on a real development project in your org. Visit blitzy.com and press book demo to learn how Blitzy transforms your SDLC from AI-assisted to AI-native. That's blitzy.com. Today's episode is brought to you by Vanta. In today's business landscape, businesses can't just claim security, they have to prove it. Achieving compliance with a framework like SOC 2, ISO 27001, HIPAA, GDPR, and more is how businesses can demonstrate strong security practices. The problem is that navigating security and compliance is time-consuming and complicated.

10:08It can take months of work and use up valuable time and resources. Vanta makes it easy and faster by automating compliance across 35-plus frameworks. It gets you audit-ready in weeks instead of months and saves you up to 85 % of associated costs. In fact, a recent IDC white paper found that Vanta customers achieve$535 ,000 per year in benefits, and the platform pays for itself in just three months. The proof is in the numbers. More than 10 ,000 global companies trust Vanta. For a limited time, listeners get$1 ,000 off at vanta.com slash nlw. That's v-a-n-t-a dot

10:44Next up in our big ideas is one that we actually have done a full primer show on, context engineering. For a bit of a definition, we once again turn to Langchain's Harrison Chase, who writes, Context engineering has been an increasingly popular term used to describe a lot of the system building that AI engineers do. But what is it exactly? The definition I like, context engineering is building dynamic systems to provide the right information and tools in the right format, such that the LLM can plausibly accomplish the task. Now, as he points out, this is not a new concept per se, but it is changing the way that we think about what our job is vis-a-vis LLMs.

11:18Whereas before many of us were just focused on how do we prompt it in the right way, now we're increasingly thinking about how do we give it access to the right data and the right information or the right tool set to go do what we need. Shopify's Toby Lutke writes, I really like the term context engineering. It describes the core skill better, the art of providing all the context for the task to be plausibly solvable by the LLM. Now, one of the things that I pointed out in my primer episode is that context engineering is itself a banner term, which includes both a dimension for non-technical people for whom it's useful to think about context engineering, even in how they prompt new set of general consumer agents like ChatGPT agent.

11:55But it's also increasingly an AI engineering discipline. There's this excellent public repo on GitHub called Awesome Context Engineering that has tons and tons of information with background, implementation and challenges, components, techniques, and architectures, and a ton of additional materials. Also, the more that people dig into this, the more interesting challenges and opportunities they're discovering. For example, in this presentation from a recent context engineering meetup, they discussed context poisoning, such as Gemini playing Pokemon hallucinated an item and then tried to reuse it.

12:38Now, going in deep on all of these is beyond the scope of this show, but the point is now that we are thinking about context engineering as a thing, in quotes, it's helping us better plan out and understand how to do it well and how to avoid common traps. HeyNita101 sums it up. Prompt engineering is for hobby projects. Context engineering is for production. Prompts are what you write in a chat box. Context engineering is what powers real AI systems. In serious LLM apps, it's not just about giving an instruction. You have to build the entire world around the model. Task setup, examples, documents, history, tool output, state, memory, compression, multimodal context.

13:12Too little and it underperforms. Too much and it breaks, slows, or costs too much. Getting this just right is not a prompt. It's a system design challenge. Next big theme, which relates to ambient agents and just in general increased agentic capacity, is tiny teams' big salaries or big cash compensation for startups. Now, it is basically gospel in Silicon Valley that when you are building a startup, you want to find missionaries, not mercenaries. And that means people who want the upside of your equity, not high cash comp, right? You're never going to compete with other bigger, more mature companies on cash.

13:46You have to give them bigger chunks of equity because they're all in on the dream. However, that truth is increasingly being questioned. Controversial founder Roy Lee from Cluely writes, A startup truth I disagree with, don't pay high cash comp. We now pay$250k to$350k base for designers and$300k to$1 million base for software engineers. In a thread explaining why, first Roy takes on the concern that this doesn't scale. You can't pay 100 people$600k a year. But Roy says, you don't need 100 people anymore. Companies grow faster than ever with less people than ever. You can hit 10 million ARR with less than 20 people in one year if you hire right and prioritize well.

14:21Speed over scale today. Consideration two, they should want to work here for our mission. Be for real, bro. You're a random B2B SaaS with zero traction and you're seven pivots deep into your mission. I don't even believe early Airbnb employees were loyal to the mission, but they might have been loyal to the culture. Roy continues, risk reward has to make sense if you want 110 % performance. Founders will earn 10 to 100x what founding engineers do if the company succeeds. You need to de-risk early employees. The worst cope is if you pay too much, people will be lazy because they need time to go spend the money.

14:50The best companies are built by the best people. He concludes, There are very few real winners in startups and winning companies are always a direct function of how strong the early team is. If you want to win, you just have to be the best at everything, including comp. A ton of people jumped in to basically say, yeah, I agree with this. Swix pointed out that this is a part of the tiny team's playbook. Others pointed out that there are fundamental problems with the way that stock compensation works at startups. And as if to demonstrate this point, we recently had the whole hullabaloo around Windsurf, where it seemed like a huge portion of the people who had helped build the company weren't going to get anything in the aqua hire by Google.

15:25Now, it would be one thing if that was just a Windsurf issue, but we've seen this over and over again. There is a new pattern emerging where big tech firms and acquirers are willing to pay an extraordinary premium for top founders and a very small handful of researchers and engineers, but then basically leave everyone else behind. What people have pointed out is that the problem with this isn't just the single one-off time that people get screwed. It's that it's going to fundamentally change how people think about working at startups in general. Legendary VC Vinod Khosla writes, Windsurf and others are really bad examples of founders leaving their teams behind and not even sharing the proceeds with their team.

15:57I definitely would not work with their founders next time. The problem is that it's not about those founders. It's about every other founder in the future where people get more and more skeptical. And so the point is when you're thinking about this big idea, it's clearly not just being driven by AI. although the patterns of acquisition around AI companies are exacerbating it, there's something bigger going on here. The smaller and more concentrated teams can be to achieve huge results, the more that we'll think differently about how they get compensated in the beginning. Next big idea is pay-per-crawl.

16:26Now, this comes directly from Cloudflare. At the beginning of July, they wrote, Publishers, we see you. Cloudflare just launched pay-per-crawl to put control over your content back where it belongs. Now, crawling is more transparent and controlled by default, creating a better web ecosystem for creators like you. This is about real content independence. And the idea here is very simple. The new program allows content owners, i.e. the people who own websites, to charge AI crawlers for access. In their announcement post they wrote, Many publishers, content creators, and website owners currently feel like they have a binary choice.

16:57Either leave the front door wide open for AI to consume everything they create, or create their own walled garden. But what if there was another way? After hundreds of conversations, we heard a consistent desire for a third path. They'd like to allow AI crawlers to access their content, but they'd like to get compensated. and thus was born the paper crawl system. Greg Eisenberg wrote,

17:43Now Greg is viewing the very optimistic side of this. He writes, What valuable data do you have sitting around? Internal company docs, customer conversations, industry insights, process knowledge. All of it can now generate revenue from AI companies who desperately need training data. The winners will be content businesses that move fast. Imagine owning a cooking blog with 10 years of recipes. That's training data gold for AI food companies. Early adopters will set market rates before competition drives price down. SEO consultant Bill Hartzer said, not so fast. He writes, Cloudflare is rolling out Paper Crawl, a new feature that allows websites to charge for crawlers like search engine and AIs to index their content.

18:16At first glance, it sounds like a win for publishers. But for the overwhelming majority of websites, it's a traffic trap disguised as a revenue stream. Cloudflare powers about 20 % of the web, so any new feature they introduce carries weight. But realistically, this one will only matter to the top 1 ,000 or so sites in the Majestic Million, those elite destinations with thousands of referring subnets. If you're not in that top 0.1%, Papercrawl probably won't make you a dime. In fact, it could do real damage to your site's visibility. Now again, beyond the scope of this show to go into a full debate, But this big idea is, I think, emblematic of an even bigger change, which is that the fundamental agreements and social contracts of the web, based around the search paradigm, are totally up for grabs right now.

18:56What the future of generative engine optimization or pay-per-crawl actually looks like, I don't think anybody really knows. All we know for sure is that the web of tomorrow will not be exactly like the web of today. Now, for the next idea, we're turning once again back to Greg Eisenberg to discuss the transition from user experience to agent experience, or UX to AX. The idea here is pretty simple. We've designed software so far for human users. Increasingly, we're going to have to design it for AI agent users. Greg writes, there's a quiet shift happening in how we design software. We're moving from UX to AX.

19:29Traditional UX is screen-centric. You tab a button, product reacts, job done. Every session starts from zero. Designers pre-plan every path with hard-coded flows. Users fill out forms and dropdowns because the product remembers nothing about you. Success equals fewer clicks and faster flows. Trust equals interface looks clean so it must work. Agentic experience is relationship-centric. The agent keeps track of ongoing goals, nudges next steps, improves over time. You're never starting over. The system plans its own path. It senses, infers, chooses actions the designer didn't script. Context is learned, not asked.

20:01Preferences, patterns, even team norms are remembered. Success equals earned trust and compounding value. Metrics shift to retention, satisfaction with decisions how much autonomy you hand over. Greg argues most apps will eventually work this way. Your email client will learn your writing style and priorities. Your design tool will remember your brand guidelines and suggest layouts. Your CRM will track relationship patterns and recommend next moves. Now, Greg is talking about a specific version of this, but the way that people are thinking about this is much broader. A great example is e-commerce.

20:28At the most recent Amazon Prime Day, Gen AI traffic was up 3 ,300%. And while that was still a tiny fraction of overall traffic, it shows just how much growth there is in new ways that people actually get to their shopping experiences. Pretty soon as agents come online, it's not just going to be that people are going to e-commerce sites from different places like chatbots, but instead that agents are actually doing the shopping for them. Those will require totally new interactive paradigms that are going to need to be designed. Two more quick ones before we get out of here. The first is the financialization of compute.

20:59This one comes out of the White House AI Action plan once again highlighted by SWIX. Sean writes, buried in the AI action plan is an endorsement that the U.S. compute market will financialize with spot and forward contracts. One of the most consistent themes with the Latent Space podcast GPU infrastructure and NeoCloud market coverage is that the status quo of three-year lock-in long-term contracts with hyperscalers is causing unsustainable market volatility and inefficiency, not just in GPU prices and the rise and fall of startup fortunes, but also inefficiency in ideas and resources for open AI and research.

21:30Now the U.S. government is fully behind this movement, and most importantly, they demonstrate that they get it. Now for a little bit more background, in April, Aiton Space did a deep dive into this, featuring Evan Conrad of SF Compute. They write, It should not be normal for the prices of one of the world's most important resources right now to swing from$8 to$1 per hour based on drastically inelastic demand and supply curves, from three-year lock-in contracts to stupendously competitive over-ordering dynamics for NVIDIA allocations, especially with increasing baseline compute needed for even the simplest academic ML research and for new AI startups getting off the ground.

22:02The ultimate end state of where all this is going is GPUs that trade like other perishable staple commodities of the world. Oil, soybeans, milk. Because the contracts and markets are so well established, the price swings are also not nearly as drastic and people can also start hedging and managing the risk of one of the biggest costs of their business. Just like we have risk managed commodities risks of all other sorts for centuries. So basically the idea here is that the financialization of commodities allows for better price discovery and more efficient markets, and that needs to come to compute as well.