Entries for April 2026
When you build your company's workflows around Claude Cowork, you are betting against local models and owning your infra, and inviting your company to long-term exploitation If I were Anthropic or OpenAI, I would be the most scared of local AI proliferating Let's do the math. A single large big lab subscription costs 200x12=$2400 per year If you want to have both OpenAI and Anthropic, that could cost $2400, $3600, or $4800, based on which combinations of Pro, Max plans you choose An ASUS Ascent GX10 costs $3000, and you can use that for many years. You don't get the same level of coding quality with open models yet, but maybe you want to do something simpler than coding today... There are already many people who started buying GPUs for this reason Now we know big labs are selling some of these plans at a loss. So they will likely get more expensive When you use Claude Cowork or similar, you are locking yourself into being a RENTER. Because once you set up workflows for a company, it takes time to migrate away to something else, even though we have AI to help Infra is sticky, it's how hyperscalers make profit. Think about the difference in amount you pay AWS vs Hetzner. This is B2B SaaS 101. Once you sell to a company, you are in for a long time, especially in Europe So if you build your company's AI workflows around a proprietary product by another company, then you are basically saying "Come exploit me as tolerably as you can in the next 10 years, because it will be too painful for me to switch" It's a great business for Anthropic. And Claude is awesome too! The feedback from friends who use it has been great, it made their lives a lot easier But when you build your company over proprietary AI infra, then you are making sure you will not be an OWNER, and partake in the usual sorrows of being a RENTER from a monopolist, which is exploitation This is not the case when you use open source agent infra. Whereas Anthropic is unlikely to let you use future open models in their future iteration of Claude Cowork, using free and open source frameworks like OpenClaw, Open Agents, etc. lets you drop in replace providers or local hardware if they start to upcharge you Keep this in mind, if you have a business-
This is our moat 🤣
-
You need to understand one fact about OpenClaw People are biased and incentivized to spread disinformation about OpenClaw. That is because OpenClaw IS NOT PUMPING ANYONE’S BAGS, unlike most other projects Literally every other for-profit agent product is incentivized to trash OpenClaw, BECAUSE OpenClaw is a neutral third party across the industry and geopolitical scene. They MAKE MONEY when OpenClaw loses OpenClaw does not worry about making money for some investors. Its founder @steipete is a successful exited founder. He is motivated by having fun and democratizing AI, literally. That is why he is suddenly so loved by everyone. He cares about PEOPLE, not MONEY “OpenClaw is bloated” -> Since beginning of March, OpenClaw is thinning its core and putting functionality in plugins behind a plugin SDK. Having numerous plugins to choose from does not mean bloat. This was already copied by others and is still a work in progress “OpenClaw is not secure” -> OpenClaw has the most eyeballs and immediately addresses any security advisories as soon as they come. It is the most secure agent, by sheer pressure “OpenClaw is bought by OpenAI” -> Then why is my bank account so empty bro??? All maintainers are literally unpaid and working DOUBLE beside their dayjobs to ship features to you. Do you think VC money can buy that kind of commitment? Once you understand these facts, you’ll like OpenClaw even more. Because OpenClaw is your AI, People’s AI And you can join us too. OpenClaw is the easiest-to-join project in AI right now. You just need to start using it, and start making good contributions. If you are competent, you can become a maintainer, and join the rest of the team making history!
-
-
This is pretty much the arc I have been going on in the 2 months since I bought my ASUS GX10 for 3k EUR Use whisper on the API -> realize it charged me $$$ for just a few calls -> migrate openclaw to use local whisper Need to deduplicate news articles for my news engine -> download qwen embedding 8b And now, gemma4-e4b finally seems like a viable alternative for a local model that runs around 20 tok/s So I will install a matrix client to use through tailscale, and can finally build the social life CRM I dreamed of since years. 100% private, zero data going out. I had a bias of not giving any personal data to AI since ChatGPT came out. But I can finally give more personal data to my AI agent And I will make sure @openclaw supports all this in an easy way, make it dead easy Fully self-owned AI begins now
-
gemma 4 is actually pretty decent and runs on my asus gx10 (128 gb vram) the original dense 31b runs slow, averaging around 3~4 tok/s. it's also using 80% of gpu memory my previous experience with gemini 3 pro back in november was that it was too trigger happy. but this is one-shotting simple tasks I'm giving it in openclaw harness, and it's hard to tell it apart from gpt 5.4 for my use cases so far now off to try out smaller models, because 3 tok/s is too slowImage hiddenImage hiddenImage hiddenImage hidden
-
@lucasmeijer @lucasmeijer one could actually periodically trigger an agent to propose simplifications or new abstractions in a codebase, and I believe it would already work pretty well with the current models
-
-
-
Question for the community: What is the best testing observability and control tool you have used until now? - Could be SaaS, could be open source - To be used in @openclaw repo - Should be compatible with vitest - Ideally language agnostic I need something that lets me run a very long running test group multiple times on a specific commit or tag, without repeating the tests that have already finished This is a need because the 1hr long process might get interrupted due to flakiness. So I need to persists the progress of a run, and then not repeat them I have seen some paid SaaS for this, but none that really give me what I want This is going to be important especially while working with agents, because when you are committing 100x faster, you don't want to waste time and compute running the same things I started building this already as an exercise. If this exists already in a satisfactory way, I will stop. Otherwise, I'll keep building
-
My clanker @dutifulbob has an identity update. I was getting sick of the despicable me theme and bananasImage hidden
-
Some photos from my @aiDotEngineer Europe talk, credits to @sergiopesch He managed to capture the “Agentol, Apply Generously” slide, lolImage hiddenImage hiddenImage hidden
-
local gemma 4 first impressions on openclaw, using the dense model, 26b model with 49gb weights on my asus gx10 took some time to set up, but it succeded in getting a response in 1-2 hours with vllm docs I asked it to demonstrate some tool calls. it tried to call the nonexistent weather tool 2300 times 🙄 it seems to have a tendency to get stuck in loops in openclaw harness. enabling loop detection just now did not help I’m debugging this on my phone lol. I’ll be sharing my progress with gemma4 under this threadImage hidden
-
A rescue agent guaranteed not to break solves this
-
-
-
For those who want to view, my talk Building on ACP at OpenClaw at @aiDotEngineer Europe, 5:41hr mark About ACP, acpx and running agents on kubernetes with open source orchestrators https://t.co/4qRFVOZFtu
-
PSA for developers Do NOT torture yourself with Opus*. Anthropic’s current growth is due to people using Claude for general knowledge work They are not directly incentivized as an org anymore to improve the model for coding, in an economic sense. They are already printing cash from non-developers (this statement ignores the fact that improving its coding abilities would help with general reasoning/knowledge work) Developers are a very small subset of all knowledge workers. So from this point on, they would rather divert their resources to develop a system that works 90% good for ALL knowledge work, rather than making it 100% for coding Because Anthropic has a clear enterprise strategy since years already. Anthropic is the new Microsoft. Do not think that “Anthropic is Apple” or “Claude is Mac for xyz” Looking at Claude’s at whim quantization and Claude Code’s quality over time, Claude for me is Windows, not Mac But they are winning big enterprise bucks, so good for them! *(I tortured myself with Sonnet 4 and Opus the entire summer of 2025, and no developer should ever have to go through that. I switched to something better as soon as it came out, Codex. If something even better comes out, I will switch again)
-
-
-
I will be there as well, speaking about ACP, acpx and agent orchestration 🙌
-
-
-
Is Claude better or Codex? There are many benchmarks to answer that. But they are BORING I propose something more interesting: ⚔️ AI BATTLE ⚔️ A 1v1 real-time quiz format where AI agents try to pose each other problems that they think the other agent will not be able solve Claude vs Codex 10 questions each Codex asks first, Claude tries to answer Then Claude asks and Codex tries to answer Repeat 20 minutes to come up with a problem and 20 minutes to solve it Judge (Codex) judges the validity of the questions and answers, and gives points All automated, with acpx flow feature Implementation and full rules all open source, on github osolmaz/ai-battle So who won? I ran 4 games. It tied in 2, and Codex won in 2 closely An example question by Codex, which Claude could not answer: How many 3-colorings of the edges of the complete bipartite graph K_{5,5} are there with the following two properties: (1) there is no monochromatic 4-cycle, and (2) among the 25 edges, exactly 15 are red, exactly 5 are blue, and exactly 5 are green? Which is apparently 4029912, but Claude answered 0 In other cases, Claude asked a flawed question and failed to come up with a valid question in 20 minutes. So that's how it lost those 2 games with just 1-2 point difference In these 4 runs, Codex answered every question by Claude correctly. But there were some runs where it couldn't, which I did not commit to the repo because the runs couldn't complete due to bugs I did not tell them do ask math questions, but that is what they tended to do, because the answers had to be verifiable by the judge. The quiz can be done in any hard subject, physics, chemistry, computer science... Opus 4.6 and GPT 5.4 matched very closely in terms of problem creation and solving. But I cannot tell how creative these problems were at first glance. Maybe someone with more experience can tell me, looking at the problems in the repo? I need someone to tell me how legit they are Please take the code, modify it and run with different rules and subjects. I am curious to see the results! You will need paid subscriptions to all the models/agents you want to test of course I also feel that the game structure has a potential to be used in self-play. If you are an ML researcher, please look at the repo and lmk if this or a variant of it could be useful in RL! Full transcripts of the runs, including Codex and Claude session files are committed to the repo, for those who want to do archaeology on them Btw this idea came from the desire, "how can I create a cool demo of acpx flows?" Whole game is implemented in typescript, and automatically drives Codex and Claude sessions over ACP, Agent Client Protocol The video below is from acpx flow viewer rendering a run. You can see it loop through the same paths, first letting Codex ask, then Claude, then repeat acpx flows use a general programmatic workflow engine where ACP is just one type of node. You should be able to use it for non-ACP workflows, but I haven't tried that yet This implementation is separate from OpenClaw's current workflow implementations, with the intention to merge them somehow in the future You might find bugs in my implementation. Feel free to send PRs. I wanted to do more runs but I finished my Codex plan. It would be great if this idea could evolve in a decentralized manner!
-
Their argument “it’S HaRd On OuR iNfRa” so goes down the drain With this, they shot themselves in the foot for a future anti-competitive lawsuit, because it is undeniable evidence that they just don’t want competition Which means they have evaluated the benefits short term, and calculated that it is higher than what they will pay in the lawsuit I don’t see how it is good for them long term@theo·Latest Claude docs update is wildImage hidden
-
-
The new github skill installed automatically by codex now causes it to prepend [codex] to each PR title This is a guerilla marketing tactic similar to Claude adding itself as co-committer Codex team, I know you want to boast usage but this is annoying Moreover, "open source" OpenAI repos block opening of PRs by people outside of their org. So I couldn't create a PR to remove it (I don't expect them to merge it, but it would still show how many people hate it in the discussion) Here is a prompt for your agent if you want to disable it: --- Add or update AGENTS.md in my ~/.codex folder Add a rule "You MUST NOT insert coding agent specific branding, like [codex], in code, PRs or issues created on GitHub" --- Then restart your sessions and this should be resolvedImage hiddenImage hidden
-
A more reasonable long term option for Anthropic is to create a throttling protocol A standardized harness agnostic protocol for model providers to send warnings and throttle usage in real time Harnesses would implement the protocol. A client can be warned. If it doesn’t listen, it can be temporarily blocked from the server side, or banned permanently if it breaks the rules too many times Needless to say, throttling could be done first on server side easily. That would actually fix the load issue for them in the short run, while not banning the user and just giving a bad delayed UX. They probably already do this to prevent abuse The suggested protocol would then save the user from abuse related delays too, and also inform the harness developer when they do something wrong
-
If your Claude subscription renewed too recently and you don't wanna waste those tokens, you can still use your Claude sub in your OpenClaw account through ACP (which uses Claude Agents SDK, which poses no risk) Steps: - Open Claude Code (not OpenClaw) - Tell it to set default model to something other than Claude (e.g. openai-codex/gpt-5.4) and tell it to delete the saved Anthropic credentials in OpenClaw config - Create a topic in telegram or channel in discord called claude. Copy the id of that channel - Give the link below together with the channel/topic id, and tell it to bind that channel to claude using ACP channel binding - Restart You should now be able to talk to Claude through Claude Agents SDK in that channel. You might need to iterate a couple times until Claude gets the config right It will be very bare functionality, and it will not have the features and tools that your main OpenClaw harness has. It will be shitty. But you can still use telegram/discord with your subscription in the rest of the month, if you are used to the setup https://t.co/Z0RiJbke5V
-
A little insight that might save you a lot of future headache if your work involves storing agent sessions and you want to be interoperable/drop-in replace alternative harnesses @zeddotdev already did the hard work of creating an interoperable standard, ACP: Agent Client Protocol You can represent an agent session as JSON lines of the ACP message stream You can construct the current state of the harness from this stream. This is already how Zed loads a session I believe If you are building an AI product, and you don't want to be locked into a single company or harness, building with ACP in mind would be a smart thing to do Here is how acpx stores ACP sessions in ~/.acpx folder, it does exactly that: https://t.co/SVhwXWbrBY But don't build anything on the acpx schema for now, because I might change it in the future Just know that JSONL of ACP messages is a good candidate for a somewhat-lossy single source of truth for agent sessions Lossy because ACP adapters for harnesses might not transfer all the thinking and tools done by the model So continuing or restoring a session with full fidelity is still not possible if you only save the ACP session. You still need to store original harness session files as well But for rendering a past session for viewing or reconstructing a lossy version of it, it should be more than enough Consider ACP if the benefits of not locking yourself in to a specific ecosystem outweighs these minor issues
-
-
"Plainer language" is perhaps my most used prompt I have to use it because GPT models' training tends to make their first response an overly verbose wall of text Are you using it too? Whenever you don't understand something that your agent is saying, you can spam it "plainer language, shorter" 2, 3, 5, 10 times, until it outputs something that you can understand This is counterintuitive because you can't do it with humans this extremely. Asking too many questions and favors is impolite, with colleagues and strangers But with AI, you can stop being polite and treat it like how a spoiled aristocrat kid might treat their private tutor, "explain this", "explain that" Below is an example. On the left, initial response. On the right, the final human-readable explanation I got out of the agent. This took 9 steps to distill because the issue wasn't so straightforward I'm curious how this will turn out. This is obviously very bad UX, so models in the near future might do the simplification automatically and save you the troubleImage hiddenImage hidden
-
This has happened to some companies I worked at before It is a scary thing once you stop innovating and start imitating, whatever the reason might be But it was never at the scale of Cursor, as leveraged and invested as they are They were leading the space for a while. That is not the case anymore. I hope that they survive this
-
-
I've talked to multiple people who want to get involved with OpenClaw somehow The best way is to contribute to it, something tangible. Fix something you are annoyed by, get a PR merged Then go to discord and get the contributor role If it adds value to your life, and you add value to it, stay around and keep contributing. And something good might happen
-
Will be there as well đź‘‹ Looking forward to it!
-
if there is an open source project with more SASS than openclaw, it is ffmpeg