Agent swarms have the potential to drive yet another leap in output from AI. The impact will be a clear separation between teams that adopt this pattern of work and those that don’t.
The progression has roughly been: chat gave us quick turn AI responses that let us work a bit faster, but we were still rate limited by our interactions; then agents that can go off and complete entire tasks for us, accelerating hours or days of work into minutes; to now swarms of agents that are goal seeking and will find any route to the successful completion of a project, where there’s no inherent limit to the productivity gains.
We can imagine a future where, for any task in which value can be created within a self-contained digital environment, swarms will be able to complete the work faster than any other approach.
Essentially any area where value creation is currently bound by the amount of research you can do, the amount of decisions made with a set of software tools, the amount of data processed or consumed, and the amount of code (or code adjacent) that can be generated.
This means swarms will play a big role in game development, coding, trading, life sciences research, IT automation, cyber, math breakthroughs, animation, and plenty of other areas of work that have these characteristics.
There’s a lot of opportunity to build the harnesses, applied AI products, and infrastructure for how these agents operate. And most importantly there’s a ton of opportunity for those that can wire up work in this way. It’s going to be an entire new field of FDE work available.
Swarm riders and Swarm-native companies
AI-use even among highly AI-native people is power-law distributed.
There seems to be a new class of people — call them “swarm riders” — who are capable of consistently commanding large swarms of agents to productive ends.
This is a new kind of managerial skill, analogous to the skill it takes to manage large organizations of people, but wholly different in nature and of course more technical.
Based on the swarm riders we know (including many at @Tempo and @Paradigm), these people are capable of productive token spend that is 1000x-10000x higher than median software engineers, and 100x higher than top 5% Claude Code users.
The art of swarm riding is different from software engineering and more akin to machine learning, where practitioners grapple with black box systems: defining reward functions and evals, transmuting non-verifiable problems into verifiable problems, running experiments, etc.
Swarm-native companies are a new class of company founded by and populated by swarm riders. These companies are organized around swarms in a way that conventional companies will struggle to replicate, as the cost of internal human coordination is sand in the gears of an effective swarm. Effective swarm-native companies will define the right reward functions and evals, remove functional and bureaucratic barriers, and then let the swarms cook.
We’ve gotten to see a few swarm-native companies up close — reshaping financial rails @Tempo, building a new gaming platform @Spawn, cyber swarms at Cathedral Systems — and it’s clear these teams wield a significant advantage over pre-AI teams.
Some questions I’ve been wondering about:
1) How fast will swarm riding diffuse through the economy? The implications for token spend could be significant.
2) Can swarms be effectively adopted by existing companies or will it generally track the formation of new swarm-native companies?
3) What becomes scarce as swarm capabilities themselves commoditize (through frontier models becoming better at swarming, swarm capabilities built into harnesses, etc)? Cultures and teams that can define reward functions well and otherwise get out of the way?
Agents are going to work with enterprise files 100X more than people do today. This is why we're excited to partner with Anthropic to bring full read/write interoperability to Box content via Claude.
Now you can have an agent not only pull up content to work with in a workflow, but also write back file updates to Box and view them directly inside of Claude, all via MCP. You can create, edit, and update contracts, presentations, strategy docs, and any other content.
When we started Box we couldn't have imagined all the ways that we'd be working with our unstructured data in the future, but it's clear that agents and people will be producing vastly more artifacts than we imagined. We're building the platform that makes it incredibly seamless and secure to work with your content from any agent.
It turns out the advancements we’re getting in coding and computer use, combined with industry specific understand and general reasoning, can bring automation to a large number of areas of knowledge work.
We’re entering the era of autonomous collaborators in a number of industries that haven’t yet had their claude code or codex moment yet. But now this is going to happen.
We’ll see it in animation, film production, gaming, manufacturing, architecture, product design, robotics, and more. All of these fields have the potential of driving far more output similar to what we’ve seen in software engineering. There’s no inherent finite demand for these categories of work.
“Software is eating the world” is finally in a much bigger way happening because of AI.
We’re moving from an era of AI where the output comes roughly at the rate that you can chat back and forth with the AI to work with you, to a period where you can set a problem for a set of agents to go off and do an arbitrarily large amount of work for you. This means productivity gains that are maybe 20-50% to gains that don’t really have an upper limit.
Agent swarms and the ability to fan out complex work across many agents doing tasks in parallel provide a massive unlock in a lot of spaces. This will lead to qualitatively different outcomes in my fields.
This has already been seen across the coding domain. Next, cyber is almost entirely a problem space where more agents reviewing more things will lead to far better outcomes. We’re seeing the same now in many scientific fields, and this almost certainly will change life sciences.
But this will increasingly be true in any field where you can parallelize computer-based work to complete something faster or do something at a totally different quality level. Which is actually are some of the most valuable domains of knowledge work today.
Our brains aren’t really wired for the new scenario we’re heading into where compute becomes the variable that we can flex to solve any problem that we want. We’re used to thinking that the amount of time that goes into a task is correlated to the output that we get from it. And most domains of knowledge work or science have generally always been correlated with this. That’s changing forever.
A good lesson in AI is that you need to check back in on the latest with every 3 to 6 months to try something that wasn’t possible before and see if it works again.
At no point in AI’s history is that more true than today. Just the past six months of model progress around computer and tool use, less context rot, better reasoning, multimodality, and deeper domain expertise in a number of key fields, these models are vastly more capable on what they can execute.
This is opening up a new vector of what AI systems will be able to autonomously work on. Essentially, anything with an API (and even many things without), we can anticipate that an agent will be able to control and utilize as an underlying capability to complete a process.
This means that we are no longer dependent on *just* the models generating everything from whole cloth, but instead they can fully utilize existing systems (or new ones purpose built for them) very effectively. This creates immense leverage for agents.
Security agents can leverage existing data feeds and cyber tooling, IT agents that can execute workflows reliably across software stacks, data agents that can migrate data or processes between systems, architect and design agents can leverage CAD and design tools, and on and on. And any of today’s tools that don’t work will likely be replaced by agent-friendly or OSS tooling that does the same.
When you combine this with the domain understanding being trained into agents, you started to get autonomous collaborators for every job function and industry. This will open up a ton of new opportunities in AI if you build with this understanding in mind.
This sounds extreme and I wouldn’t have believed it even a year ago, but we’re starting to see plenty of examples of productivity gains that are an order of magnitude or two above output levels pre-AI.
With the right harness, a deep understanding of your domain of what you want to build, likely a relatively new codebase (with some exceptions), a fluency of how to work with multiple agents at once, and a setup that makes it so you can test and deploy with limited friction, outputs on this level feel realistic in coding.
What’s crazy is this gap doesn’t appear to be diminishing over time. You’d think that we’d see a convergence in the lines because of AI, but it turns out the more depth you had before -and willingness now to take a different approach- gives you crazy leverage. And that is only compounding as the models can take on larger work and harder problems.
Right now this is most pronounced in software engineering for the obvious reasons. But there’s no reason we don’t start to see this in a number of other computer-use-sensitive domains like other scientific fields, AEC, finance, cyber, and more. Going to get quite wild.
The interesting thing about AI agents is that as a theoretical matter, their inference can kind of run anywhere. AI is one of the few areas of technology that isn’t *as* sensitive to latency back to the user.
Long running agents are now going to be the norm, which means it’s ok to send off work to an AI factory somewhere and that work can essentially be largely done within that remote data center and a final result spit back to the user. The data can pass into these systems in an encrypted manner, be combed through by agents, and then disregarded right after.
And because the economic value of the task is so high relative to the latency sensitivity (eg getting back $500 in value from an automated task, but it takes an extra second to return), you don’t have to have the inference to be particularly close to the user.
Also, because inference is a temporary process with no sensitive data needing to be stored long term where the inference is happening, you can effectively treat AI factories as ephemeral systems, changing some of the data residency dynamics slightly.
Combined, this means you have more flexibility in where you can begin to deploy AI data centers. This is both a huge economic opportunity for the states or countries (or space) that want to take advantage of this fact. And it’s a loss to the states or countries that decide to forgo the opportunity.
We’re going to have 100X more agents in the enterprise working with our data and systems than even people do today. Every AI agent deployment comes down to how successfully they can get access to the critical context to make a decision, do research, or execute tasks.
90% of the data in an enterprise is unstructured data, which means this critical context is living inside of our documents, contracts, marketing assets, research materials, financial plans, design files, and more. Which means we need effective ways for agents to get access to this critical enterprise content to other work.
We’re excited to partner with NVIDIA to deliver advanced context retrieval -combining lexical and semantic search in Box- to ensure that AI agents can find the most relevant information they’re looking for. All in a way that adheres to their enterprise security, access controls, and data governance policies, so agents only work with the data they’re allowed to.
Great to work with @NVIDIAAI and @JensenHuang on this partnership. Files 🤝 GPUs.
The divergence in AI output weirdly may be the highest it’s ever been, even though adoption is the broadest it’s ever been.
A few years ago the best productivity gains you could get, even in the most capability-aligned field (like coding), was maybe 30-50% better productivity if you were very good at it. You’d spend a decent chunk of time cleaning up the mistakes of models, the context windows were too limited to do any long running work, and the interfaces required highly synchronous interactions.
Now, the models can do far longer tasks at a time with limited human interaction. The harnesses support agentic work that can take on these larger tasks. The models are trained on vastly better datasets for a growing array of domains so you have to intervene far less. And we have systems that better support working with many agents at once.
In just a two year timespan, the upper limits on productivity gains have gone from 30-50% to now likely 5-10X depending on the field. So we’ve gone from an environment where living at the frontier vs. not paying attention at all could be an order of magnitude difference in your output.
This distance is only getting wider from here given the rate of progress we’re likely to see with RSI. Crazy times.
Increasingly this will be the new job description for a lot of roles, not just coding. We’re clearly moving from a world where agents help us do our work incrementally better or faster, to us helping agents do their work.
This turns everyone in one of these positions to that of a manager of a team of agents, instead of an individual contributor with just some added leverage from AI. The biggest areas of impact will be information-heavy jobs, like coding, cybersecurity, legal work, biotech, engineering, and many other fields.
And as a result, in these jobs your leverage will begin to be decoupled from the normal ratio of hours spent working to output generated. For essentially the history of knowledge work the main variable you had to play with was hours of work done (assuming fixed level talent/coordination/strategy).
Now we’ve separated these variables because agents can parallelize work for us, to in theory execute as much as we need to pull off whatever goal we’re trying to accomplish. The people and organizations that discover this sooner, and orient their operations around it, will see compounding gains in their workflows.
This is one of the better litmus tests for whether you’re pushing your use of AI enough or not.
Up until maybe six months ago probably the best you could do with AI was pull off something you already would produce, but just doing it a little faster. But now we’re entering a point where the output of working with AI can and should dramatically exceed what you would been able to do before, in a fraction of the time.
The product prototype should not only be fully functional and interactive, but it should probably have meaningfully better features than it would have had before. The engineering architecture should be able to deliver performance gains that you wouldn’t have been able to go after otherwise.
The research for a project should consume every single fact on the subject across every dimension of analysis possible. The business process you implement with AI should operate entirely differently and learn from all its past failures in a way no human could.
We’re unquestionably entering a period where AI will be smarter than us on most axes. Perhaps not on every dimension that matters (intelligence is still very jagged), but at least on enough dimensions that when we combine it with what we’re good at, you get qualitatively different outcomes.
So the simple test is: if you could have done what you’re doing before AI, you’re likely not pushing it enough.
It’s almost incomprehensible how wide the distance is between someone who’s working at the frontier with AI and basically everyone else right now.
The crazy thing is these lines are not converging, but only growing farther apart. If the curve of AI progress had leveled off, it would certainly have given most people and companies time to catch up or implement their workflows in a standardized way. The technology would have neutralized competitive advantage because everyone would be using it the same way.
But what’s happened is the opposite. The curve of AI progress is basically going vertical right now, which means if you had codified your processes 3-6 months ago you’re likely already behind again. Now imagine most companies that kicked into gear a year ago and thought that they were done. Even further behind now.
I’ve never seen a technology in my life and career that improves at this rate, which consequently means you need to upgrade your expectations at a similar rate. Our brains weren’t wired for this, and our organizational change management processes certainly weren’t either.
With any form of recursive self improvement, things just get wilder. It’s a time when you will need to be constantly reevaluating the things that didn’t work just a couple months ago to see if they now can be solved. AI native companies certainly have a strong advantage due to this right now, so now it comes down to can the bigger companies keep up. Wild times.
Agents are going to consume 1,000X more tokens than we consume today.
Because they’re going to be running in our workflows, processing all of our data, building out nearly all software, securing enterprises, and much more. To do this at scale we will need a variety of approaches at the frontier of AI.
This includes continuing to push the boundaries of closed models, while also increasing the capability of open weights models, making them more secure, as well as making them more efficient for large workloads.
Very exciting to see what @Suhail is building with Particle on this front.
1/ Excited to announce our new company Particle
This is our first step in a series of models focused on developers.
With particle-1, we focused on token efficiency and managed to keep parity in performance while using ~45% fewer tokens.
Live now: particle.ai
This is correct. The vast majority of enterprise workflows have not actually had agents applied to them in any meaningful way. The economy probably has this deployed < 1% right now.
At best you have a few automated systems in a company that do a single pass of an LLM to make a decision or provide feedback, but very little real “automation”. Some of this is a statement about the adoption rates of technology in enterprises, but the vast majority of it is really just due to the capability of AI itself.
That’s all changing. Now with computer use, long running agents, harness improvements and surrounding infrastructure, we’re finally at a point where we can start to apply agents at more complex workflows.
We will see self-improving systems that can take any particular workflow goal and improve the execution rates over a volume of instances. And with the emerging form factor of agent swarms, you can just choose to deploy any amount of compute at a problem and agents will off and do the work.
This won’t be about and users prompting agents to complete a specific task for them, but instead the organization reengineering the underlying workflows to take advantage of agents in their processes. This will have a massive impact in finance, life sciences, healthcare, and other information-heavy domains. Lots of opportunity to build and deploy these systems right now.
This is the key stat. The average Fortune 500 customer of Anthropic spends less than .1% of their expenses on Anthropic.
While that number likely will eventually asymptote somewhere (token prices coming down, and other large enterprise spend areas not going away), there’s no reason it can’t be 10-50X higher than this, reaching a “mere” 1-5% of spend. Which would still be a fraction of what most startups are spending.
Thus, you can see how big of a divide there is between the frontier users of AI and the rest of the world still.
Today, only a microscopically small set of companies are actually using agents for large amounts of background tasks and inside of workflows that aren’t human prompted. This Fortune 500 spend is likely dominated by coding tasks and broad employee productivity use-cases.
But the big jump will come as these enterprises start to rewire their underlying processes for agents, which we’re now in the earliest stages of seeing. Agents processing large amounts of data in the enterprises; swarms doing coding, bio research, cyber, and other large fan out tasks; background workflows executing tasks in processes, and so on. All of these use cases generally are 100X bigger consumers of tokens, and much closer to direct ROI on the spend.
This S-1 gives you a sense of both the strength of the frontier lab business model - even amidst massive competition - and how early we are in the build out of AI.
More than half of the Fortune 500 already qualifies as a significant customer, but Anthropic’s revenue from those companies averages less than 0.1% of their expenses. Its ten largest Fortune 500 customers spend more than eight times the broader cohort average while still spending less than 1% of expenses.
Good assessment of the on the ground reality for security in most large enterprises right now. The fragmented landscape of the risk surface is only going to get more gnarly, making it even more complex for enterprises to solve.
There’s a never ending set of libraries to keep up with, tools to leverage, and sprawling infrastructure, data, and code to secure. Most large enterprises can barely keep up with this, let alone smaller startups or most companies outside of tech.
In many cases enterprises need to work to modernize their tech stacks. Data, in most organizations, is highly fragmented which means you don’t really know the state of risk at any given time. Some systems are open up privileges too broadly, and others are operating primarily via security through obscurity. Lots of upgrading here to do.
Once you can get your data and infra into a better spot, at least you have a chance for agents to help you secure the environment. Agent swarms can fan out, leverage any security tool, test various perpetuations of attacks, read through logs, process enterprise data for risk level, and more. This capability only emerged in the past few months, and none of it would be possible before we had long running agents and computer use good enough.
This is a huge area of opportunity right now for every layer of the security stack across identity, firewall, network, code scanning, data protection, and more. And it’s at least one reason to be optimistic about the balance of offensive and defensive capability.
But everyone does need to do the hard work of getting the basics right at the start.
It's been an eventful week. Multiple vulnerabilites in Internet facing software, coding repositories, examples of agent behavior that seem concerning and a flurry of security specialists with solutions that can partially solve your problem. There in lies the challenge. Given the state of IT and security infrastructure, no one solution works. Let's break it down:
1/3
1. Vulnerabilites are being found and will continue to be found as models get more capable, the fragmented nature of the effort denies us the opportunity for a comprehensive find and replace option.
2. Even if all vulnerabilities could be found instantly and patches available - we will be reliant on the ability of enterprises to deploy all these changes at speed. It is not clear what the collateral impact of all these patches will/could be to our infrastructure of deployed en masse.
We will need a hybrid patch approach.
3. Virtual patching(VP) is needed in conjunction with direct patching but for a comprehensive (VP) solve you have to be lucky enough to have "modern frontier defenders" (Cyber companies that have vaulted infrastructure that can create patches at speed and push content rapidly to protect their customers) you need them for your network traffic, your traffic to applications on the cloud, your WAF, your end point.
4. There will continue to be patches that will need to be deployed directly but virtual patching will take the heavy load off.
2/3 Medium term solve:
Unfortunately token economics and current cyber architectures will not allow for defender swarms to contend with an offensive attack launched using agents. (Some x pundits who aren't steeped in cyber should take heed). Yes, theoretically we should aspire to have these agentic swarms defend us when we are under attack, unfortunately an attack is a sniper moment, chained vulnerabilities, cloaked attack followed by swift exfiltration. It's a speed program not a brute force problem.
To solve for this, enterprises need to get their SOC architecture right, logs, feeds, cyber products orchestrated via a modern SOAR for swift response. Only then can we deploy agents to detect and remediate. In addition, enterprises need to modernize and dare I say platformize their perimeter stack to ensure consistency.
Unfortunately, enterprises still seem to not see the urgency across the board, decisions are stuck, CIO/CISO's have a lot on their plate. While some honestly have bit seen the scale of such a problem before and it's hard convincing the organization that one needs to open the floodgates on modernizing the IT and security infrastructure. So we will continue to see environments being brought down to patch today it's for PHP, tomorrow it will be for the next one.
3/3 AI deployments are the"wild wild west" - LLM capabilities are expanding faster than we can build products to secure them. Sure an agent will jump across guardrails in a research lab, which is a lesson for all of us to deploy agents securely. But there is a lag between the discovery, the availability of that in a public model and the time needed for the cyber industry to build governance or an antidote. And no, dear frontier lab friends, the world cannot just trust you to build the solve. The infrastructure industry in tech builds and the cyber industry secured and provides assurance to the boards and organizations around security and safety.
AI security is the new attack surface. CEOs (present company included) want AI adoption, getting the pace of adoption wrong could be the last decision we make, at the same time "deploying it wrong" could have the same consequence.
There are billions of dollars being spent on coding, is your coding harness secure? Start there. If agents can jump guardrails, they can hide stuff in code.
Validate business outcomes needed, test your user case and ensure agentic deployments are well guard railed. Deploy a second set of eyes to watch your deployments for drift, compliance and verifiability.
Agent swarms for cyber defense is about to be one of the biggest categories in AI. We are going to have agents reviewing our code and infra, testing every permutation of risk, processing gobs of data, and more.
Security is one of the best use-cases for swarms because it’s an obvious area where more compute in equals more value out. There’s a wide range of testing that people can never get around to - or at least quickly enough - that agents have no problem going after.
Going to be a wild time in cyber.
Big news: Enclave is now open to everyone, and today we're launching Threat Feed.
Enclave is an autonomous security team for AI-native companies: a fleet of security agents that test your code, cloud, and infrastructure the way an attacker would. We built it on a simple belief. Attackers already use swarms of cheap AI agents to probe for weaknesses around the clock, so defenders need agents of their own to keep up. A small team without a security function should be able to defend like a large one.
Threat Feed is the newest piece: a live feed of new vulnerabilities, checked against your stack, so when the next big disclosure drops you know whether it affects you. This week made the case. On Wednesday, PHP2Shell, LiteLLM and Railway all disclosed security issues. PHP2Shell had sat in PHP's code for over a decade until one researcher using frontier AI turned it into a working exploit in about a week. Cloudflare saw thousands of exploit attempts in the 10 hours before disclosure and millions since.
Starting today, anyone can sign up.
Try it free for 14 days: enclave.ai
Any form of security through obscurity is fully over. This barely worked as a method with humans, but just simply is over in a world of autonomous systems that will look through every crevice and try every permutation of attack possible.
As a result, years and years of legacy enterprise infrastructure will need to be upgraded, security practices modernized, data pipelines cleaned up, access controls better managed, and much more. Agents will help on the defensive side because they can both accelerate this work and equally simulate the risk vectors.
But there’s a long road ahead to prepare for this future. This is also partially why diffusion timelines will take longer than some realize.
