I moved web search in my agents to @ExaAILabs over the weekend, and wow, the difference is very noticeable.
Compared to native web search and Brave Search API, the quality of the results and the usefulness of the context coming back to the agents is a major upgrade. Better retrieval changes the quality of the whole agent workflow.
83
My goodness, the last few @openclaw updates have been so painful. Hours debugging why the restart does not work, and then config is broken and incompatible with previous versions. Doctor fix can fix the config, restart does not work.
92
For full agent autonomy, we need strong controls
The deeper I go working with agents 🤖, the more I keep coming back to the same point.
If we want agents to consistently deliver on tasks as autonomously as possible, we need to implement strict, well-defined controls.
If you want agents to operate with real autonomy, you need a clear control layer around them. Boundaries. Permission models. Auditability. Systems that make it easy to understand what happened when something goes wrong.
That has become very real to me as I have worked through this in practice. The more productive and continuous you want the autonomy to be, the more important the control layer becomes.
And you can see that in the most popular coding agent today: Anthropic ‘s Claude Code. These control mechanisms are built into Claude Code harness. In my latest post, I wrote about the Claude Code leak and what it teaches us about this topic. It is very interesting to see how the harness handles permissions, tracks denials, and creates accountability around agent behavior.
If you are an engineering leader working on agents, I think this is worth understanding more deeply.
Link 👉 antonio.md/posts/claude-code…
70
Some very important updates from both @OpenAI and @AnthropicAI and getting closer to AWS.
For enterprise teams, having these options available through @awscloud and Bedrock can remove a lot of friction. Many companies already have strong enterprise data platforms, security controls, and infrastructure relationships built around AWS. That means they can move faster, expand access to these models more easily, and avoid having to negotiate separate contracts just to get started. They can also leverage their EDPs. We did that when we started using Claude Code. It was just faster to use Claude code via Bedrock.
I think this will help more companies bring OpenAI and Anthropic into their stack at the same time. That matters because the pace of innovation is so high right now, and teams need the flexibility to use the right LLM platform for the right job.
This is also a smart move for OpenAI and Anthropic. As demand keeps growing, and compute is a limiting factor, meeting customers through Bedrock gives them an option to offload the underlying compute burden to AWS.
I think this is a great outcome for everyone.
🔗 Claude on AWS: aws.amazon.com/claude-platfo…
🔗 Andy Jassy’s post: xitter.cf/ajassy/status/20488060…
Very interesting announcement from OpenAI this morning. We’re excited to make OpenAI's models available directly to customers on Bedrock in the coming weeks, alongside the upcoming Stateful Runtime Environment. With this, builders will have even more choice to pick the right model for the right job. More details at our AWS event in San Francisco tomorrow.
70
Really impressed with @ouraring AI support experience.
I got my issue resolved quickly in a chat flow with clear predefined answers, and it gathered the right context without asking me to go hunt for anything.
This is what good AI support looks like. Not canned AI answers that go nowhere.
1
255
For a long time, we have been talking about the models.
I have been reading a lot about the lessons from the Claude Code leak and breaking them down into a few areas. The first one is the agent harness.
One of the most interesting lessons from the source code leak, at least in my opinion, is that the real product is the harness around the model.
The model call itself is one part of the agentic system. If you already have an LLM provider that satisfies you, what matters in practice is everything else around it: prompt assembly, memory, tool routing, retries, permissions, session continuity, and all the operational work required to make agents reliable.
I have seen the same thing while building my own agents using @openclaw and other harnesses. Recently, after @AnthropicAI changed the policy around using the Claude Max subscription, I switched my agents to GPT 5.4 through my @OpenAI Pro subscription. The harness stayed the same. The configuration stayed the same. And that is exactly the point: what makes the agents useful is not just the model. It is the surrounding harness 🤖
Here is what I wrote about it in part one 👉 antonio.md/posts/claude-code…
115
Very good article from @hwchase17 about the agent harness and memory achitecture.
⭐️ Saturday mornings are perfect for learning new things. 💡
One of the biggest problems in agent engineering is managing context and memory, especially in long, complex sessions.
All agents use some form of compaction. It helps keep the session alive, but it is lossy. Important details get flattened, and execution quality gets worse over time.
All the frontier LLM companies are working on this, but most of it is happening behind closed doors. That is why it is interesting to watch what 🦞 OpenClaw is doing with its new Dreaming feature. I think it is a real step forward. It treats memory as something that needs curation, promotion, and hygiene, not just bigger context windows or gigantic memory files.
I wrote more about it here 👉 antonio.md/posts/why-agent-m…
1
37
Not all agent work is equal 🤖🦖🕵️
The key question is not just what an agent can do. It is whether the human operating it can verify the output with confidence.
Some agent work is easy to check. Some looks good on the surface but is much harder to judge. In my latest post, I build on Bian's framing and add another lens that matters a lot in practice: impact radius.
Worth being explicit about whether you are building with agents. face, but is much harder to judge. In my latest post, I explained this framing and added another lens that matters a lot in practice: the impact radius of a mistake.
Worth being explicit about if you are building with agents.
I wrote more about it here: antonio.md/posts/not-all-age…
31
Now that @AnthropicAI is blocking using the Claude subscription with 3rd party harnesses, I moved all my Agents to use @OpenAI GPT 5.4. What happens to my agents was very interesting.
Is Agent personality a thing?
I wrote more about that here 👉 antonio.md/posts/model-behav…
🤖 Made with AI
1
2
74
Great new releases from @OpenAI today at the #OpenAIDevDay event. ChatKit, AppsSDK, AgentKit, tons of updates to Codex and New APIs
1
228
Just saw one of the new @Waymo cars in SF, finally ditching the Jaguars for a custom self driving design.
141
I am finally able to test the new @Tesla FSD 13.2.2 and it is really a major leap when compared to where the software was just 12 months ago. It is unbelievable how much more confident and agile this version is. Wow.
1
238
🍎 @Apple finds itself in a predicament. I’ve been using Apple Intelligence with iOS 18.1 📱 for quite some time now, and while some features are promising, they lack the compelling power to truly delight users. ✨ Writing Tools is the standout feature, but its capabilities are very limited. AI photo editing 📸 falls short of the sophistication offered by Google Photos. Siri has improved, especially with better contextual understanding, but after using @OpenAI ‘s advanced voice mode, I expect that Siri should be able to have a long, open-ended conversation with me. Instead, when you try to chat about anything beyond basic questions, Siri presents you with a @Google Search result 🔍 and then the conversation stops. That is really bad— I don’t want a search result with links to sites; now I want the answer. This partnership with Google is undoubtedly lucrative 💰 for Apple, but it also hinders their ability to innovate and deliver a truly exceptional user experience.
#AI #Apple #OpenAI #Siri
1
2
599