I Tested OpenClaw Against Model Churn. Here's What Survived.

Añadido: 2026-05-10

[00:00:00]In April 2026, OpenClaw grew up. The teenager not only got their hands on the keys to the car, they actually grew a prefrontal cortex and can make responsible decisions. Now, that's what it feels like if you are an OpenClaw user that started in January and you're now in April. Because OpenClaw has been adding features at a break neck pace that basically amount to adding a responsible thinking adult brain into the system. What's happening is that Peter and team are adding the ability to execute complex multi-step orchestrated workflows across many different tasks in OpenClaw. The beauty of OpenClaw was always that it was a simple architecture that was extensible that could do many many things. That's why people love it so much. But as models get better, you want to be in a place where you can orchestrate more complex runtime environments, more complex agent environments, more complex agentic tasks. And that is the heart of what OpenClaw released. But there's a problem. And the problem is pretty simple. When you do harder work, you have a hard time assigning all of that work to one brain, right? To one LLM.

[00:01:07]And so what I want to give you today is how to think about these April results thoughtfully and strategically for your claw. You need to think about where does it make sense to put one LLM in versus another. How do you not be dependent on LLM? And yes, I'm going to give you specific perspective there. And really this is the critical piece. What do you need to build so that you are not stuck when Anthropic makes another change which they did this month. We'll talk about it. When OpenAI makes a change, which they did this month, we'll talk about it. Both related to Open Claw. I want you to have your own claw that does its own work. And you need the ability to control the core workflow loop without having to depend on one LLM to do it. Because if you can swap LLMs, you can do so much more with your claw.

[00:01:58]There are so many more use cases. And yes, if you're wondering, what are those use cases? I have a claw. I've been wondering. We will talk about that. We will get into that in this video. And yes, there's a whole build here. If cloud people are builders, we have a whole build to improve the durability of your workflow so it can survive this model war that is happening in April over the OpenClaw agent brain. So, let's hop into it. After all of the progress for OpenClaw in April 2026, OpenClaw is starting to look less like a viral agent demo and more like a real runtime that gets real work done. And once that happens, the model fight that people report on gets a lot more interesting because the model is no longer the whole work product. It's a brain inside a much larger work loop with a lot more impact.

[00:02:40]So I want to separate very clearly three things that are getting mashed together in April around openclaw. First, open claw itself got more mature. Second, the model layer around it got much more contested. And third, once the runtime can swap brains, memory becomes the strategic layer. That third point is the one I think people are underestimating.

[00:02:59]If the claw can run many brains, which it can, the memory should not live inside any one of those brains and should be built to adjust to the intelligence you want to apply to a particular workflow. So if you paid attention to OpenClaw a month ago and are kind of checking back in now, the simple update is, wow, there were a lot of releases and that is true. OpenClaw shipped at a pace that would be exhausting for a normal product team and almost absurd for an open source agent framework. There were task updates, memory updates, provider updates, channel updates, code and automation updates. The release notes alone feel like a product team sprinting while the rest of the market argues about model access. But release velocity is not really the whole point here. The point is that the shape of the product changed with all these releases. A month ago, it was still pretty easy to describe OpenClaw as a viral open- source agent framework. It gave a model access to your computer and your files and your browser and your apps and your messaging services. It was very powerful. It was a bit messy. It was risky in the way all serious agent systems are risky and it was compelling because it made the model feel like it had hands. That description is still true. It's just incomplete now.

[00:04:03]Open claw is starting to clarify what kind of thing it actually is. The claw is becoming a particular species of crab or lobster. It's not just a chatbot rapper. It's not just a claw launcher.

[00:04:16]It's not just a terminal toy for people who want to see how much power they can hand to an AI assistant before something weird happens. Open Claw is becoming an action layer for agents. More specifically, it's becoming a runtime abstraction for serious agentic work.

[00:04:30]That sounds abstract. So, let me make it very plain. A chatbot is a place where you ask for help. An agent runtime is a place where work happens. And when you are starting to experiment with agents, the bar is how demanding is your agent runtime at quality so you can do serious work. And that's the progress we see with OpenClaw in April. And I'll get into the details. A serious agent runtime has state. It has tools. It has permissions. It has retries. It has handoffs. It has context that needs to survive more than one prompt. It has a user who needs to know what happened, what failed, what changed, and what ought to happen next. That is a big shift, right? Open claw is moving from can I make the agent do something to can I build a durable work loop once and route different models through it to get a bunch of different work done. That is a way bigger idea than original open claw and it's a wonderful example of how building the right primitives allows very fast progress in the age of AI. And so hats off to Peter for building an extensible primitive here. The clearest evidence for all this is the boring stuff. And I do mean that as a compliment. When agent products are immature, the exciting features get all the attention. A model opens a browser.

[00:05:37]A model sends a message. A model buys a car. That was one of the stories that I covered right out of the gate with OpenGL. That stage matters. You need the demo moment. But the next stage is where the product either becomes infrastructure or just is a party trick.

[00:05:49]And infrastructure announces itself with really boring words, right? Tasks, cues, histories, checkpoints, visible delivery, scoped memory, provider manifests, permission profiles, retry behaviors, tool boundaries. These are not the words that make a launch go viral. They are the words that decide whether anyone can build on the system for serious work. Open clause task flow work is a good example. The docs now describe task flow as the orchestration layer above background tasks. It manages durable multi-step flows with their own state and revision tracking while the individual tasks stay a unit of detached work on their own. And that really matters because a task you can inspect and route and cancel and recover and deliver back to the right channel is way way different from a chat response. It's way way different from the original Open Claw vision where you didn't have that idea that you would want to manage multiple tasks like that with your lobster. A web hook triggered workflow is different from a user manually typing please do this and that is again something that serious work needs. A sub aent that can run its own session and report back reliably when it's done is different from one big threat. So the product maturity layer is starting to come together and it's easy to underrate because it doesn't look like a single feature. It's a dozen different updates, dozens of updates that make the agent less fragile. And that is exactly the point. The magic demo is the model did something at all and that was February.

[00:07:11]The mature product is we are actually building real work here. Memory is part of this same story. Early agent memory of course was mostly a novelty. The bot remembers your name. The bot remembers that you like TypeScript. The bot remembers a preference and inserts it later in a way that feels delightful or maybe slightly creepy. That gets attention. But serious work needs a very disciplined memory model. If an agent is operating on a repo and reviewing PRs and triaging incidents or maintaining customer feedback, memory can't just be a pile of things the model said or that you said to the model. You need to know where memory comes from. Was it observed from a real source? Was it confirmed by a user? Is it stale? Is it scoped?

[00:07:48]Should it be retrieved automatically?

[00:07:49]Does it tied to a particular LLM or not?

[00:07:52]Those questions sound boring and fussy until you use an agent for real work.

[00:07:55]And then they become the difference between useful continuity and just a bunch of sludge that accumulates in memory. Open clause memory direction with memory wiki active memory and a lot of providence rich recall points toward a mental model of memory that's helpful here. Memory is not just personalization. Right? Memory is operational context and that changes what the whole product is. If the agent remembers what happened last time, it can do way more than answer, right? It continues. It reviews against repo conventions. implements against meeting decisions and it starts the next incident with the last incident in view.

[00:08:28]And that's the second piece of our story here. It's not just about actions, it's about continuity. Channels are the third big piece and this is another place where the boring details matter. Open claw being available across Slack and Telegram and Discord and WhatsApp and Teams and Matrix and Fu and other surfaces is not just a distribution flex. The channel is an important part of the runtime. Different channels have different rules. Threading is different.

[00:08:51]Mentions are different. File limits are different. Bot permissions are different. An agent that finishes a task but never replies visibly is broken in one way. An agent that replies in the wrong thread is broken in another. So mature channel behavior isn't glamorous, but it's a really important surface to handle. And this is where the open claw story starts to look less like an agent that kind of developed in a chat app and more like a runtime that happens to have a bunch of humanfacing surfaces. The human might be in Slack. The agent might be doing code work in GitHub. A sub agent might be analyzing logs. And there may be multiple models involved here, right? A stronger model might handle the hard stuff while a cheaper model does a first pass classification of intent.

[00:09:27]From the user's perspective, the work just needs to come back right in the right place with the right amount of detail done at the right time. That is a larger conception of what an agent can solve and open claw is maturing enough to get there. So that's the key big story in April. OpenClaw got more mature, but the market drama entered at the same time because while OpenClaw was becoming much more serious, the model layer underneath it got really, really contested. Anthropic's April move was extremely disliked by the developer community. I can't paint it any other way. Claw subscriptions were of course never designed to power always on thirdparty agents at scale. That is the basic anthropic position and I kind of get it. Agents aren't normal chat users.

[00:10:09]They run longer. They retry. They call tools. They carry more context. They sit in loops. You can't price them the same way, right? They create intermediate work that a human never sees. That's also tokens. They can turn a flat rate consumer subscription into infrastructure for some derivative product. And from Anthropic's point of view, that is money losing, right? If Claude is being used as infrastructure, Anthropic wants it paid for like infrastructure. Use the API, buy extra usage. You need to protect the margins that make Claude a sustainable user experience. And that is rational, but it was deeply unpopular. And it's also a strategic choice and I don't want to hide those other pieces. If claud is being used as infrastructure, Enthropic wants it paid for like infrastructure.

[00:10:49]Use the API, buy extra usage, stay inside official products, protect enthropics capacity, protect enthropics margins, protect the direct claer experience. Seems really rational from anthropic's point of view. It's also a very intentional choice that we shouldn't run away from. Anthropic is effectively saying Claude is powerful enough to be an agent brain, but Anthropic wants to control the terms under which the brain is used. It doesn't make Claude less useful, but it changes the default architecture for a lot of claw builders who started out with Claude. Claude becomes a very premium metered component to Open Claw, not the cheap always on substrate for background loops. It's a different design assumption and frankly it is costing Anthropic with a developer community even if it's the correct decision for them from a token usage perspective. Remember we have to view a lot of anthropics actions in the past month as a function of hyperrowth leading to compute constraints and they are making tough choices all across the board to handle that bind and I have a lot of empathy for them because that's not an easy place to be in. This is one of those tough choices and yes it was unpopular. It doesn't mean that they had a lot of other options. I I expect they knew it was unpopular and they made the call. Anyway, OpenAI has more online compute and they are taking a different posture. Codex is already an Aentic product. OpenAI's help docs now make CEX part of the chat GPT subscription across all paid tiers and Open Claw's provider docs describe a Codex OOTH route alongside direct API usage. That matters because open AI is essentially looking at open claw and seeing distribution within the plan. In fact, Sam Ultman said that very explicitly on May 1st calling out that open claw is now just flat available under chat GPT paid plants and that is basically the opposite of what anthropic chose to do in April. Think of it from OpenAI's point of view. If they have the compute, then if OpenClaw users route more work to chat GPT and codeex, it reinforces OpenAI's agent strategy. It makes codeex more central and it makes open AAI infrastructure feel like a natural place for open agent workflows to land. And yes, the Peter Steinberger piece matters here too. When the creator of OpenClaw is working at OpenAI and OpenAI is making codeex more available as a subscriptionbacked agent surface, the power dynamic changes. Enthropic had the model many early users loved. OpenAI now has codecs, subscription access, and a strong incentive to make these workflows feel native on its infrastructure. But I don't think the responsible version of this story is as simple as open AAI wins and entropic loses. I think the larger story is that there are many ways to run your agent. And that's a principle that's worth defending. Open clause should not depend on one provider subscription policies to work. It should be able to run on 5.5, claude API, Gemini, DeepSeek, Open Router, O Lama, LM Studio, Gemma for whatever comes next. The point of the runtime abstraction is that the action layer can stay stable while the brain changes. And that is the real strategic idea. And that is why Google's Gemma 4 is not a side quest here. Google launched Gemma 4 under Apache 2.0. And the positioning is very explicit. These are open models built for advanced reasoning, agentic workflows, and ondevice use. The Google developer post is even more direct.

[00:14:00]Gemma 4 is meant to bring Agentic skills to the edge, including multi-step planning, autonomous action, offline code generation, and multimodal processing on device. that matters because it gives builders a credible local branch of the overall runtime tree to play with. Not every agent step deserves frontier model pricing and Google is recognizing that here. And this is where the model conversation gets super practical. The old argument was which model is best. The better argument is which model should handle this step. Use a local Gemma class model for cheap background classification for duplicate detection for low-risk triage.

[00:14:34]Use GPT 5.5 through codecs for hard implementation and complex repo work.

[00:14:38]Use cloud API when the judgment, the writing style, or maybe the architectural reasoning is worth the metered cost. Use cheaper hosted models for bulk summarization or formatting.

[00:14:47]Your workflow should be sophisticated enough to survive those routing decisions. And that is the aha moment that I see a lot of builders hitting in April. You can build one work loop once for open claw. And you can stop treating model choice as a permanent architectural decision. Don't be fooled by the arguments about claw versus open air. That's not the point. And this is the headline for builders. The practical unlock is not simply that open claw can use different models. A model drop down.

[00:15:14]Oh, fine. It's convenient. But if you are swapping your entire runtime brain that is a strategic shift you need to plan for. How do you design workflows that outlive a model, a subscription plan, a provider policy, a context window? That's what I mean by a durable workflow. A durable workflow has a job to do, a place to run, memory of what happened before, and enough structure that the underlying model can change without destroying the workflow. The model still matters, of course, sometimes it matters more, but it is no longer the product surface. The workflow has its own identity. It has inputs, outputs, permissions, tools, state, review steps, a human-facing channel, a failure mode, memory. The model becomes the reasoning engine inside a much larger operating loop. And as you swap the model, you can do different things with that general purpose loop. Let's imagine you're handling a repo, not an agent that can code in the abstract. I mean a workflow that watches GitHub issues and pull requests over time. It triages incoming work, compares new issues against past fixes, knows which files are risky, and remembers the tests that usually catch regressions. The useful knowledge is not all in the prompt. It's in the history of the codebase. old review comments, prior bugs, deployment failures, style preferences, architectural decisions, and the uh we tried this and it broke staging lessons that every real team tends to accumulate. If that memory lives inside just a chat transcript, obviously your workflow isn't going to work for very long. If it lives inside one provider's product, the workflow is locked to that provider. If it lives outside the model, the runtime can call whichever model is right for the step and still behave like the same operator.

[00:16:49]That is the idea of a durable workflow.

[00:16:51]A local model can classify the issue and gather context. GPT 5.5 and codeex can make a patch. A review model can inspect the diff. Claude can write a high judgment architecture pass if the change is sensitive. You can use whatever model you want here, right? I'm not saying you have to use those models. The user doesn't have to care which brain handled which substep. The user cares that the repo operator understood the job and left the team with something reviewed.

[00:17:14]Code review is a super clean way to talk about this, but I want to call out there's lots of examples where we have durable workflows that are non-technical as well. Like if you are trying to do a scheduled review that has multiple layers of your email inbox, which I'm naming email because it's like the number one most popular use for openclaw across nontechnical users. Then that's multiple jobs, right? You have to segregate out sensitive emails with high judgment. You have to draft replies if you're going to draft replies. You have to automatically review those replies for quality, tone, how it responds to the sender, how it handles the original message. You have to figure out how you are threading replies correctly and make sure you have a QA pass for that. You have to handle any artifacts that come with emails like attachments correctly and securely. You have lots of tasks with email that would require models.

[00:18:03]One, remembering how you like this done, two, executing the workflow and the right model getting called. and three, having a durable memory of how all of this goes together so that the next pass works. Incident response is the last example I'll call out here in the video.

[00:18:17]If you're interested in diving in more to how to use Open Claw in practice, I have net new workflow examples in the Substack, including customer feedback loops, meetings to execution, and more.

[00:18:27]For this video, I'm picking incident response because the pattern it shows spans a bunch of surfaces, and I think it illustrates, right? You have to deal with logs and dashboards and Slack and GitHub and runbooks and deployment history and customer reports and previous postmortems and a live timeline and everyone is panicking and asking what changed at the exact moment when like they're not thinking straight. A durable open claw workflow sits across all of that. And when an incident starts, it gathers context and identifies changes and compares symptoms against prior incidents and drafts the first update and suggests roll back candidates and turns the resolution into a postmortem draft. The point is not a magical model does all of this for you.

[00:19:04]No, you may want a fast model for logs, a cheap model for updates, a deep inference model for root cause, but the instant workflow shouldn't need to care at the product level which brain did which step. The same pattern is useful across a bunch of other tasks. And that brings us to the common layer underneath it all, memory. If the workflow is durable and the brain is swappable, memory can't live inside anyone brain.

[00:19:27]This is the point that connects openclaw to the open brain project. When agents were mostly chat companions, shared memory was useful. When agents become real workers, the way OpenClaw has matured, shared memory is something you have to handle. A worker needs to know what happened before, right? Project conventions, people involved, decisions made, etc. If that memory lives inside a single model product, regardless of what job you're doing with your claw, you have a lock-in problem. If it lives in a random chat transcript or only markdown files, you're going to have a retrieval problem. If it lives in the agent scratchpad, you have a continuity problem. If it lives in a userowned memory layer, you have an architecture problem. This is why I think an open brain recipe for claw makes sense. Now, look, open brain doesn't need to chase every open claw feature and it won't.

[00:20:13]But open claw users by the thousand are already wiring in open brain on their own. And it just makes sense to publish a recipe that makes that easier. It just makes the whole architecture and the whole approach very legible. And so open claw agent memory for open brain is now live in our open source GitHub repo. The job here is not to add another abstraction or to complexify your claw.

[00:20:37]The job is to make the existing pattern very easy to adopt and go through and sort of to publish what works so that thousands of users aren't having to roll their own approach here. So what does the open brain recipe for claw do? It defines how the agent retrieves before meaningful work starts for your claw specifically for serious workflows. The claw is now capable of for project context, people, decisions, prior failures, current tasks, constraints. It defines how the agent writes back for serious work for outputs, lessons, unresolved questions, source channel, model use, task ID, confidence, etc. And finally, user confirmation status. That last piece matters. Agent written memory needs providence. Was this observed from a source, inferred by a model, confirmed by a user, imported from a transcript?

[00:21:20]How do we get it? Is it valid? Can it be used as an instruction later? Without clear labels, memory gets dangerous. But when you have good labels, memory gets operational. And so, yes, I'm releasing multiple recipes here that help you realize this overall vision of open brain for open claw. I'm releasing a code review memory recipe that stores reusable lessons from PRs, a task flow worklog that records what a long-running agent attempted, what changed, what blocked it, what the next agent should know, and releasing a memory and provenence recipe that labels where the memory was observed and confirmed and imported from. And that becomes especially important in a brain swappable world, which Open Claw can now do really, really well for serious work.

[00:22:01]And this is what helps you build one loop and swap those brains in and be confident that the memory actually works so that you can actually attack different problems with the same agentic pipeline. Bad memory makes the agent confidently wrong in a way that often feels personalized. But a good memory architecture makes the agent operate continuously without making it unaccountable. And that's the line I'm looking to draw here with this update for open braids designed specifically for OpenClaw. Now, if you want more detail on these recipes and as I called out on the use cases, I have that on the Substack. Of course, you can head over to the open-sourced open brain repo and get all of the code right away. But for this video, the important point is the architecture. Open Claw is maturing into a runtime. The model layer is becoming swappable and very contested by model makers. Therefore, memory needs to be userowned. And that is really, really important as we see this OpenAI versus anthropic story continue to unfold. We need to be confident that agent builders can design for a world where models can be swapped out really easily. A serious agent workflow needs to survive model churn and pricing changes and better local models and everything else we know the future is going to throw at us because otherwise how are you going to take advantage of the serious workflow features that allow your claw to be more than a toy and be really useful? Just think about it. If channel delivery is unreliable, the agent may do work no one sees. If permissions are sloppy driven by memory, a more capable model makes the system more dangerous. If providence for memory is absent, then the memory itself is untrustworthy. All of this boring product work is what makes exciting work with your claw possible.

[00:23:40]So the postappril open claw thesis is actually pretty simple. Open claw gives agents an action layer that's really works now. Models provide a reasoning engine. Task flow gives work a durable loop. Channels are where humans interact. Memory is a continuity layer.

[00:23:54]permissions and providence are a trust layer. It's good architecture. Once you see the stack this way, the opportunity for builders changes. The opportunity is not make another shallow claw wrapper.

[00:24:02]That layer is going to get crowded out really quickly. The more interesting opportunity is to build vertical work loops on top of the runtime. Build for sales operations, build for research workflows, build for meeting follow-up, build for compliance review, build for chief of staff loops, for finance analysis, for personal knowledge maintenance. In each case, the product is not an agent. The product is the loop that is tied to that workflow. And the scarce asset is not just access to a model. The scarce asset is ownership of the memory, the tools, the permissions, the operating rhythm around the model.

[00:24:33]Open claw is becoming more useful because it is becoming less dependent on one interaction platform. Look, the labs will keep fighting over open claw's brain. That's inevitable. Frontier models are expensive to train, expensive to serve, and strategically critical. I expect open AAI, Anthropic, Google, and everyone else to make choices that reflect their incentives. Some paths will open, some paths will close, some companies will push APIs, some will push subscriptions, some will push official agent products, and some will push local models. The builder response should not be religious loyalty to any provider. It should be architecture. Build the runtime so the model can change. Build the memory so the user owns it. Build the workflow so it survives the session.

[00:25:13]Right? build routing so cheap that models can do the work they need to do.

[00:25:18]Open claw has crossed into serious work mode and once you can run multiple brains through the serious work layer that openclaw has added in April, you are in business for a bunch of new use cases, but only if the memory lives independently of that brain that is so hotly contested in Silicon Valley right now. And that is your takeaway for Open Claw in April. Stay with me. We have lots more model updates coming this week. Cheers.