AI Engineer World's Fair 2026
Every Harness Will Become A Claw
Read the talk
Every Harness Will Become a Claw
Coding assistants are growing into persistent, always-on systems with initiative and learning, but greater capability does not guarantee a lasting place in users’ lives.
From a talk by Sam Bhagwat
Before you start: Familiarity with LLM tool calls and basic coding-assistant workflows will help; no Mastra experience is required.
From a coding assistant to an autonomy spectrum
What changes when the coding assistant you use locally becomes an agent your whole team can message in Slack? The model may still generate code, but the surrounding system must do more. Sam Bhagwat approaches that expansion as co-founder and CEO of Mastra, a TypeScript agent framework, drawing on roughly eighteen months of observing agents in production.
The starting landscape contains three kinds of building material: local harnesses used for everyday coding, cloud harnesses sold as products or built internally around Slack, and open-source frameworks that supply primitives for constructing either. The move beyond an agent loop is already visible in how people access and operate these systems.
Earlier discussions centered on agents versus workflows. A broader comparison is the spectrum of driving autonomy: lane assistance, Tesla FSD, and riding in a Waymo with nobody behind the steering wheel represent different degrees of responsibility delegated to a system. Bhagwat similarly arranges LLMs, agents, harnesses, and claws along an autonomy spectrum. Its successive qualities are action, durability, always-on operation, and eventually initiative and learning. These are useful architectural distinctions rather than standardized autonomy levels.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
An agent carries state through repeated action
The first transition is from a single model response to an agent loop. The model can call tools, inspect results, retain memory, and retry a failed task. Context engineering determines what information it receives on each pass. Agent state carries the work forward instead of requiring every invocation to begin from nothing.
These capabilities belong together: a tool call lets the system act, while state lets a subsequent turn respond to what happened. A failed operation can become input to another attempt. That repeated, stateful interaction is the foundation on which a harness adds longer execution and richer control.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A harness keeps work moving
A harness adds durability and doggedness: the ability to keep pursuing a task through interruptions and over longer periods. Bhagwat describes work lasting hours or days as the intended capability, not a measured runtime result. His concrete recovery example is a connection lost during a turn: if streamed progress has been persisted, the session can resume from retained progress instead of starting over.
Planning mode and parallel subagents give that persistence a work structure. A harness can split a task and run parts concurrently. A terminal user interface, slash commands, and skills provide additional ways to direct it, while dynamically created agents avoid requiring every specialist to be defined in advance. Background batch tasks let work continue outside the foreground exchange. Automatic compaction handles a full context window, and persisted threads let a user return after disconnecting. Bhagwat uses Claude Code and Codex as familiar examples of this capability family; the list is not a version-specific guarantee for either product.
Live control changes the rhythm of interaction. Users can queue, steer, and interrupt while the system works. Bhagwat compares strict alternating turns to Civilization, where a player waits for the other civilizations to finish. Steering a running harness is closer to StarCraft or Age of Empires: new instructions can arrive while activity continues.
Permissions also acquire a longer lifetime. Approval can cover a session rather than a single tool call. Bhagwat’s joke about approving every invocation of rm -rf / exposes the danger: a broad permission can authorize destruction before there is any meaningful opportunity to reuse it. Session approval reduces repeated interruptions, but its scope has consequences.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Always-on access changes the execution architecture
The intermediate step toward a claw is an always-on harness, a shift Bhagwat says had become especially visible over the preceding few months. In a shared Slack conversation, several colleagues may give instructions to the same agent. The system must interpret those instructions using user metadata, rather than treating every message as if it came from one undifferentiated operator. Mobile access adds another entry point; an app may tunnel back to a harness running on a local machine.
Access and execution location are separate choices. A mobile interface does not necessarily mean cloud execution. When execution does move into cloud sandboxes, more resources can support more parallel subagents than one machine can accommodate.
| Arrangement | Execution location | Architectural consequence |
|---|---|---|
| Mobile tunnel | Local machine | Remote access to local execution |
| Cloud sandboxes | Distributed cloud resources | More capacity for parallel work |
The additional capacity requires a different architecture and brings the trade-offs of distributed systems. Moving to the cloud changes more than the address of the process.
The output path changes too. A coding harness may move from editing local files or a Git worktree to pushing work through a GitHub pull request. That makes the result available through the team’s shared development workflow. Adoption remains uneven: some developers still work entirely with local harnesses, while others are figuring out how cloud agents fit into their organizations.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A claw initiates contact and learns from runs
The next transition adds initiative and learning. Consider an assistant that notices an apparently urgent email and texts its user to ask whether the message is actually urgent or just spam. The user did not begin this interaction. An external feed supplied something to examine, and the agent initiated contact. A heartbeat provides another trigger: the system wakes at a defined interval and decides whether something needs attention.
A small TypeScript example makes the distinction between noticing a message and resolving it explicit. Here, an urgent-looking email becomes a proposed question; its urgency remains unconfirmed.
typescript
type Email = {
id: string;
subject: string;
urgentCandidate: boolean;
};
type PendingQuestion = {
emailId: string;
text: string;
status: "pending";
};
function reviewInbox(emails: Email[]): PendingQuestion[] {
return emails
.filter((email) => email.urgentCandidate)
.map((email) => ({
emailId: email.id,
text: `Is "${email.subject}" actually urgent, or is it spam?`,
status: "pending",
}));
}
const inbox: Email[] = [{
id: "email-42",
subject: "Urgent: confirm today's delivery",
urgentCandidate: true,
}];
const pendingQuestions = reviewInbox(inbox);
A feed listener or periodic heartbeat could invoke this review. The result is a pending question, not a sent text or a confirmed emergency. That boundary preserves the uncertainty in the assistant example.
Channels make that initiative reachable through text, WhatsApp, or Telegram. Memory can live somewhere more accessible than simple file storage, while an always-running daemon or gateway handles incoming and outgoing requests. Together, these mechanisms let the assistant maintain continuity outside a terminal session.
Learning adds a feedback path from execution traces back into future behavior. A run can produce a reusable skill, or the system might modify the code that drives it. In this discussion, continual learning means adaptation through artifacts such as skills and code; it does not establish that model weights are being trained during use. Bhagwat emphasizes that the right method remains unsettled. Automatic modification is a mechanism to explore, not evidence that every modification improves the system.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Claw capabilities become framework primitives
Projects such as OpenClaw and Hermes Agent make these capabilities concrete enough that framework users begin asking for them. Bhagwat describes demand for their features with greater power and control: builders want more than installing a claw on a machine. They want primitives they can assemble into their own agents and harnesses. That demand shapes Mastra’s direction.
The progression now has a clear shape: action becomes durable, persistent work; that work becomes always available; availability gains initiative and learning. Bhagwat gives the resulting prediction the informal name Steinberger’s law, joking that he has not asked permission to use the name: “every harness will expand until it becomes a claw.” It is his forecast about where user demands lead, not an established technical law.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Users pull harnesses into more of their lives
The pressure to expand is technological, economic, and psychological. Once a harness produces useful work, users want more ways to reach it: a Slack DM, a text message, or a request sent before bed so that a task runs overnight. Availability follows an existing desire to delegate.
Bhagwat calls the reward loop a “dopamine casino”: put in tokens and get out code—or other useful actions, since agents extend beyond programming. He credits Dex Horthy for the accompanying image. The appeal helps explain why users keep asking for broader access and greater autonomy.
But the expansion phase does not imply that every product survives. After systems become more powerful, Bhagwat expects a shakeout. To understand that second phase, the relevant constraint shifts from what a harness can do to how many services a person can keep using.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Capability competes for limited attention
Android and iOS provide the analogy. Mobile platforms made directions, ride hailing, payments, and music available on phones. Video, browsing, food ordering, and accommodation booking expanded the range further; some activities migrated from desktops, while others emerged over the decade. A platform could support many categories without users adopting many competing products in each category.
Bhagwat points to Uber and Lyft as an example of people keeping a small set of familiar services. His explanation is limited mental space: even when alternatives exist, users can remember and return to only so many products. This is a consumer analogy, not a measured market-concentration result.
Two routes can keep a service relevant:
- High value when needed. Airbnb may be used only occasionally, but accommodation can matter a great deal at the moment of booking.
- Frequent use. Uber or DoorDash can recur often enough to remain readily available in a user’s mind.
Bhagwat contrasts these with Thumbtack, which he characterizes as less frequent and less valuable in this comparison. That is his assessment of perceived usefulness, not an independent valuation of the business. His proposed retention condition is that a product needs high economic value or high frequency; lacking both, it is easy to forget.
The same fading relevance happens with a college friend after a move or a change in careers and social circles. The relationship may once have mattered, yet fewer recurring circumstances bring that person to mind. The analogy explains why initial enthusiasm alone cannot sustain a product’s place in daily life.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Build for capability growth and selective adoption
Today’s excitement comes from turning tokens into useful results. Bhagwat expects categories to emerge as that excitement matures, leaving users with room for only a limited number of claws. More capable systems still compete for finite attention.
Builders therefore face an immediate obligation to keep learning what users need. If the pace of change accelerates, understanding the field must become more frequent too; this is the motivation behind Bhagwat’s appeal for events and continued learning. An agent that lacks newly expected capabilities risks replacement by a more powerful alternative.
Capability leadership does not remove the second risk. Even a product that reaches the top of its category must survive the consolidation that follows. Bhagwat places that possible shakeout in the later 2020s. The practical tension is to keep pace with expanding autonomy while building something valuable or frequent enough that users continue making room for it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Current quickstart for building TypeScript agents and harnesses with tools, memory and scheduling.
Personal assistant project connecting models, tools and messaging channels through a Gateway, with setup and security guidance.
Nous Research's agent with persistent memory, reusable skills, scheduled tasks and a messaging gateway.
Sam Bhagwat's guide to agent tools, memory, workflows, evaluation and deployment.
Further reading
Bhagwat's companion essay connecting rapid technical change, industry events and the demand for interpretation.
Updates since the talk
Current documentation for periodic agent turns, scheduling and notification routing.
Read the complete timestamped transcript
- 0:00
[upbeat music] I'll get us started. Um, so long day of talks and, uh, how are y'all feeling? [audience cheering]
- 0:20
Cool. Yeah, good to see we still have some energy. You know? I, I know there's a lot of, like, evening events. Um, we've heard a lot about the present, and I'm gonna talk about the future.
- 0:30
Um, so my talk is called Every Harness Will Become a Claw. Um, here's a little bit about me. Um, I am the co-founder, CEO of Mastra. We are a TypeScript agent framework.
- 0:42
Um, I am also the author of a book that you may have gotten a copy of either outside or at a previous event. Um, we have seen a lot of agents running in production, um, over the last 18 months, and I'm not-- uh, I'm saying that as kind of context for, and stage setting for the thoughts and
- 1:01
ideas that I'm about to share right now. Um, and, uh, the thing that I'm gonna say is, is welcome to the harness era. What do I mean by the harness era?
- 1:12
Um, well, um, let's just talk about the types of harnesses that we see right now. We see local harnesses. Um, we use them every day, daily driving our, our coding, right?
- 1:22
Um, we see cloud harnesses. Um, these are both products that we can purchase as well as if we work in some of these, uh, companies that have built their own internal coding agents that live on Slack.
- 1:33
Um, and then of course, we have the, your friendly local, uh, open source frameworks that have some of these primitives and give you the, the tools that you need to build your own.
- 1:43
Um, and that's where we, that's where we fit in. Um, now let's talk about where we are sort of collectively as an industry, um, and how things have evolved over the last y- we'll say year to, to 18 months.
- 1:55
Um, last year at, at AI Engineer, we were talking a lot about agents. We were talking about the agent loop. We were talking about agents versus workflows. Um, so, so I wanna, you know, there's...
- 2:06
As we're thinking about, um, the agentic spectrum, I often compare it to, uh, self-driving as a spectrum, right? There are different levels of self-driving autonomy, whether that's like lane assist, whether that's Tesla S- uh, FSD, whether that's I s- I'm sitting in the back of my Waymo and there's nobody behind the steering wheel, right?
- 2:24
Um, there are various aspects to the agentic spectrum between LLMs, agents, harnesses, and claws, and I'm gonna talk about what we've seen and where we're going.
- 2:35
What makes an agent different than an LLM? Hopefully, we mostly know this, but just as a quick refresher, right? It's the agent loop, it's tool calls, it's memory, it's the ability to retry failed tasks, it's context engineering.
- 2:47
Um, Dex is a close friend and an inspiration for this, uh, one of the inspirations for this talk. Um, and it's agent state, right? These are some things that, like, hey, I'm running an agent in a loop, and I can't just do this with a one-shot call, uh, to, to an LLM, right?
- 3:03
I d- um, I've tried to make these qualities. I don't-- Not sure what the quality of taking actions is, active or something, so I just put action. But, um, you know, qualities here are starting to emerge when we move from an agent to a harness.
- 3:17
Durability and doggedness. Um, a friend of mine was referring to an agent that he was using, and he called it dogged, which I really like, and I'm, I'm taking that for this talk, right?
- 3:25
So durability, just the sheer quality of, like, being able to run not for minutes, but for hours or days. Um, you know? What, what, what, what encompasses this? Well, sometimes it's like, hey, I, you know, I, uh, lost a connection in the middle of the turn and, uh, you know, but I persisted the stream, and so now
- 3:43
I can resume from the place where I started, right? There's planning mode. We all see this in Claude Code. Um, parallel subagents being able to fan out multiple tasks at the same time.
- 3:53
Uh, we have more affordances with a TUI and slash commands. We have skills. Um, w-we don't have to define all our agents up front, but we can dynamically create them on the fly.
- 4:03
This is, you know, very powerful. Um, y-you can sp-- Uh, the, the harness can spin up background batch tasks, right? Um, it will auto compact when it runs out of the context window.
- 4:14
Uh, you know, these are all things we'll see when we use Claude Code or Codex, right? Um, it persists. It will persist threads, right? You can resume a thread once that's, you've, like, disconnected from later.
- 4:27
Um, you can queue, you can steer, you can interrupt. You're not just blocked waiting on the LLM. Hey, I take a turn, and then you take a turn. I'm playing, playing Civilization here, and I can't take a turn until all the other civilizations are playing.
- 4:38
If you know, I'm playing StarCraft, I'm playing Age of Empires. Um, right? Uh, you know, session long, um, tool approval. So it's not just like, "Yeah, I approved this specific tool call," but yeah, you can run all instances of RM, RF, forward slash, right?
- 4:52
That, that you see in the session, even though the first one will probably wipe your machine. Um, okay, so there, there's like, you know,
- 5:00
there's a, um, there's a few steps here and I'm, I'm, I'm about halfway through these, and then afterwards we're gonna talk about what it means. And this is kind of a in-between step.
- 5:11
I think this is something we've seen over the last really three months. Um, and I think we're also starting to grapple with what it means, which is this movement from a local harness to a cloud harness where the harness is always on.
- 5:26
What do I mean by a harness that is always on? Well,
- 5:31
you might be talking to it in Slack. Maybe you're talking to it in Slack along with your colleagues, right? Um, maybe you're each giving it instructions and it has to figure out how to parse that and use user metadata.
- 5:43
Uh, maybe you have a mobile app. I was just, uh, uh, you know... Maybe you have a mobile app, maybe it tunnels to your local, um, to, to your local machine.
- 5:51
Some of these, uh, some harness mobile apps do this. Um-
- 5:56
Often, like, cool, how is this running? Well, it's probably running in a cloud sandbox 'cause it's-- maybe it's running locally in your machine, you're tunneling into it, but maybe it's just running in a cloud, in the cloud somewhere, and it's got a bunch of sandboxes, which enables more parallelism.
- 6:09
You can get more, um, parallel subagents beyond what you can do on your machine. This is always a trade-off and always something you get with distributed systems, right? You can do more in the cloud than you can do locally.
- 6:20
You have more resources. It requires a different architecture. It's more powerful. Um, and then lastly, you're not creating code, you know, on your-- just on your local machine or maybe even in a Git work tree.
- 6:30
Um, you're, you're probably creating, you know, if you're writing code, you're probably creating a PR that, that pushes, right, to, to GitHub. Um, so, so, you know, there's a shift, right, from, from local harnesses to these always-on kind of like cloud harnesses.
- 6:44
We're still in the middle of this. You may have-- You may only be working with a local harness. You may have started to see cloud harnesses pop up in your, your organization, right?
- 6:54
You may be figuring out how to use them. Your teams may be figuring out how to use them. Um, and then I wanna talk about what the harness-to-claw transition is, which is imbuing these agents, imbuing these harnesses with initiative and, and learning, right?
- 7:10
What is initiative? Well, um, if you've used, let's say, a, a personal assistant, uh, a-agent, right? And that agent texts you and says, "Hey, I saw an urgent email come in.
- 7:22
Is that email actually urgent or someone like, you know, spamming you?" You know, t- but, like, like, the, the agent is listening, um, to external feed services. It has a heartbeat, which means it wakes up every, you know, defined amount of time and, um, and does something, right?
- 7:38
Uh, again, like channels, so- uh, you might be able to text it, WhatsApp it, Telegram it, where-- whatever you want. Um, you, you may persist the memory, memory in a more accessible later place than just simple sort of like file storage, right?
- 7:53
Um, you, you might-- it might have a, a daemon, it might have a, a gateway, uh, for, for sending and, and receiving incoming, outcoming requests. Uh, it often will do continual learning, right?
- 8:06
So this concept that, you know, the agent and the harness runs, and then, you know, based on the traces that it generates,
- 8:14
it, it sort of auto-improves itself. And there's different ways of doing this. You see, um, skill gen-- automatic skill generation, for example, is a common one. Um, it could modify the code driving this as well.
- 8:26
Um, we haven't figured out what the right way of doing it is yet. We're still exploring, you know, the industry is still exploring options. Um, now the reason that-- a-and, and maybe this is like our unique vantage point here, but, you know, for the last three months as a framework, we've just seen this as a fu- as
- 8:45
the future. And so we furiously looked at the, the, you know, the features that, you know, OpenClaw have, that Hermes Agent have and say-- and, and we- we've said, like: Look, you know, a lot of people, a lot of folks want these features, but they want them with power and control.
- 8:59
They don't wanna just put a, you know, a claw on a box, right? They wanna have more. And so, you know, we, we've been thinking about this because
- 9:07
we, we, you know, my-- I'm not doing my job well if I'm not giving everybody the tools that they need to build agents, to build harnesses as- with the maximum power, right?
- 9:18
Um, so, so hopefully, like, again, we-- hopefully, I've walked a little bit through the step transition with actions, durability, doggedness, always on, initiative, learning. Again, I, I think I failed in, like, making them all the right tense phrase and making them all qualities, but I hope you get the idea here, right?
- 9:36
Um, we're ascending on the agentic spectrum. Um, and what was a simple LLM eighteen or twenty-four months ago is a lot more powerful. So I've called this, uh, without sort of asking consent from Pete, but I've called this Steinberger's law, which is, I, I believe every harness will expand until it becomes a claw.
- 9:56
And, and, and that's a little bit, um,
- 9:59
uh, technological, it's a little bit economic, that's a little bit psychological. So let me walk you through the reasoning here. Um, the first thing that I, I've observed, um, that we've all observed, um, as a, as a, uh, is that harnesses tend to expand, and they expand because we want them to expand.
- 10:17
Um, we want to DM them in Slack. We want to text them and, like, start overnight, uh, tasks before bedtime. We want this dopamine casino that we get when we put in tokens and get out code, right?
- 10:31
Um, or, or whatever other actions, you know. Um, uh, agent- agents are bigger than just coding agents, but, um, we want our own dopamine casino, and this, this image is thanks to, uh, to Dex Hartie.
- 10:43
Um, but I see something else in our future, um, which is that-- a-and it's something that, like, I don't think we, we sort of talk about as much. Uh, but after this, after this phase where, where we're sort of making everything more and more powerful, um, there will be a shakeout.
- 11:07
Um, and, and let me walk you through sort of, uh, th-through my reasoning here, which is that
- 11:14
in the twenty tens, we had these platforms. We had Android, we had iOS, and all of a sudden, there were all these things we could do on our phones that we previously weren't able to do.
- 11:26
We could get directions, we could hail rides, um, we could send payments, um, we could play music. Um, other ones emerged over the course of the decade. We could watch short-form video, um, we could watch long-form documentaries, we could browse the internet, you know.
- 11:42
Some were kind of ported over from the desktop. We could browse the internet, um, again, you know. Some, you know, we could order food, right? Um, we could b-book accommodations.
- 11:52
But, but if you look at most of these kinds of categories, and there are quite a few categories, there are really only, like, one or two, you know, logos here that we use, you know.
- 12:02
Okay, how many-- Maybe, you know, we use Uber and we, we use Lyft, but, like, do-- anyone use another ride-hailing app here? You know, like-
- 12:12
A-and, and so w-when you talk to people that are smart about, like, consumer behavior, the, the reason they say that this is, is because you only really have space in your brain for, like, a limited number of things.
- 12:28
Like, y-y-if, if-- Think about something like Thumbtack. So, uh, Thumbtack didn't really serve a very high economic value. Like Airbnb, like, we only use it, Airbnb occasionally, but when we use it, we, like, really want it.
- 12:40
We really need it. You know, it's really valuable to us. Thumbtack, like, a little bit less so, right? Um, a-and then it's also, like, not frequent, right? Like, maybe U- Uber, maybe like, y-you know, something like DoorDash or Uber, people c-can use multiple times a day, right?
- 12:54
So there's, there's sort of like, it either has to be very economically valuable or it has to be very frequent. And if it's n-neither one of the two, um, we just forget about it, right?
- 13:04
It's like that, you know, college friend that, like, we haven't really talked to in years. It's not 'cause, like, they weren't important at one point in our life, but, like, there's just nothing that-- Maybe they moved to a new city or we moved to a new city or our lives, our friend groups, our careers diverged.
- 13:18
And all of a sudden, like, you know, maybe we're calling them once a, once a year or once every other year or, or whatever, and
- 13:25
y-y-y, you know, there's just nothing that makes them pertinent and brings them up in our, our, our brains.
- 13:30
And so right now we're, like, really excited because there's all this energy and excitement, and we're all excited at these harnesses that we're, like, s- you know, putting in tokens and getting out, like, useful things that we all love.
- 13:40
Um, and, and I think that in the not-so-distant future, there will be this very real shakeout and, and these categories will kind of emerge, and we'll realize that we only have space in our lives for so many of these claws.
- 13:54
They're very powerful. We, we love them very much. Um, and so I would, I would think about, um, what, what, what does that mean for you? So the, the first thing is, um, the first thing is don't get--
- 14:09
This is the reason that, like, events are, are, are important, that, like, staying up with the-- Like, if the ch-- rate of change increases three to four X, that means, you know, we need to figure out what's going on [chuckles] even more frequently.
- 14:21
That's why we're all here. Um, but if you're building an agent, make sure that it has the capabilities that your users need, because if it doesn't and if, if there's newer things that come out, like, they may just, you know, pick, pick something that's more powerful because that, that's happening very quickly.
- 14:36
Um, and then keep in mind that if you, if you, if you aren't, if you're-- the thing you're working on, um,
- 14:43
e-even if you climb up to the top of the hill, keep in mind there's going to be another wave of this, sort of these, like, w- this, this shakeout coming and, you know, probably sometime in the later twenty twenties.
- 14:55
Uh, so, um, the, the-- I'm Sam. Um, I'm the, uh, I'm the co-founder of Mastra, the, the TypeScript agent framework. I'm the author of Principles of Building AI Agents.
- 15:05
Hopefully you can get a copy of the book outside, or I've got a few here. Um, please stop by, say hi. Um, it's great to see all of you.
- 15:13
Thank you all for coming out. It's a real pleasure. Um, enjoy the rest of the conference. [audience applauding] [upbeat music]