Bill Krueger

Over the past month, some of you reported Claude Code's quality had slipped. We investigated, and published a post-mortem on the three issues we found. All are fixed in v2.1.116+ and we’ve reset usage limits for all subscribers.

@garrytan AlphaClaw doesn't auto-repair issues like this for you?

Literally can count the hours I have lost as OpenClaw decides to do brain surgery on itself and then ends up dying from an edit it did to its own config

I'm usually not one to write thought pieces without much technical depth. But here we go. Slow the fuck down. mariozechner.at/posts/2026-03-…

@cramforce @RhysSullivan OMG the Petstore API is back 💀

Imagine just-bash supports OpenAPI, MCP, and GraphQL discovery for codemode style tool calls! I hooked @RhysSullivan's excellent executor package into just-bash to do exactly that. And you get to freely handle tool approvals in your JS code.

@YousifAstar Vercel introduced a method of credential brokering to their Sandbox product that injects credentials on egressing traffic. Does that fully address this kind of attack vector or would it still be susceptible to your approach?

@enesakar @ikbear @upstash @openclaw How do you easily enable this when things like crons and maintaining websocket connections when the box sleeps while idle?

@ikbear @upstash @openclaw we can easily enable this, but not supported as of now. we wanted focus on development cases first, so not sure we should host personal agents initially..

Introducing Upstash Box - the best way to give your AI agents a computer 🎉 ◆ Secure, isolated cloud sandboxes ◆ Built-in Claude Code, Codex or OpenCode ◆ Sleeps when idle, wakes up in milliseconds

I honestly wonder if we don't skip right over $500 and go straight to $1000. "Introducing our new GIGA plan!"

It is hard to communicate how much programming has changed due to AI in the last 2 months: not gradually and over time in the "progress as usual" way, but specifically this last December. There are a number of asterisks but imo coding agents basically didn’t work before December and basically work since - the models have significantly higher quality, long-term coherence and tenacity and they can power through large and long tasks, well past enough that it is extremely disruptive to the default programming workflow. Just to give an example, over the weekend I was building a local video analysis dashboard for the cameras of my home so I wrote: “Here is the local IP and username/password of my DGX Spark. Log in, set up ssh keys, set up vLLM, download and bench Qwen3-VL, set up a server endpoint to inference videos, a basic web ui dashboard, test everything, set it up with systemd, record memory notes for yourself and write up a markdown report for me”. The agent went off for ~30 minutes, ran into multiple issues, researched solutions online, resolved them one by one, wrote the code, tested it, debugged it, set up the services, and came back with the report and it was just done. I didn’t touch anything. All of this could easily have been a weekend project just 3 months ago but today it’s something you kick off and forget about for 30 minutes. As a result, programming is becoming unrecognizable. You’re not typing computer code into an editor like the way things were since computers were invented, that era is over. You're spinning up AI agents, giving them tasks *in English* and managing and reviewing their work in parallel. The biggest prize is in figuring out how you can keep ascending the layers of abstraction to set up long-running orchestrator Claws with all of the right tools, memory and instructions that productively manage multiple parallel Code instances for you. The leverage achievable via top tier "agentic engineering" feels very high right now. It’s not perfect, it needs high-level direction, judgement, taste, oversight, iteration and hints and ideas. It works a lot better in some scenarios than others (e.g. especially for tasks that are well-specified and where you can verify/test functionality). The key is to build intuition to decompose the task just right to hand off the parts that work and help out around the edges. But imo, this is nowhere near "business as usual" time in software.

Seeing the backlash from @NotionHQ users around the pricing of their new Custom Agents feature, it is clear the average person has no idea how much AI inference really costs. It is difficult to build in this space when you cannot heavily subsidize that cost like the big labs do.

@cramforce This is a great post. It really cements the mental model I have been developing around how to safely provide users with a powerful agentic harness that abstracts the complexity of something like OpenClaw while protecting both our secrets AND theirs.

This is an important conversation to be had: What is the right architecture to running agents in production? My posits: - You can trust agent harnesses, you cannot trust agents - The agent harness, the agent, and the code it generates should not run in the same security context Today this separation is not typically enforced and this will lead to major security issues in the future. The bright side: Existing agent abstractions do make it relatively easy to port to a more secure architecture. More concrete in the blog post ↓

@vercel_dev Awesome, can't wait to check this out!

You can now write chat bot logic once, and deploy across Slack, GitHub, Teams, and Discord. We're open-sourcing a new Chat SDK, a unified TypeScript library to build for many chat platforms with a single codebase. Now available in public beta. vercel.com/changelog/chat…

@Shpigford @replysocial Ouch 😬 I am a bit surprised they don't allow the Pro tier to continue programmatically replying to posts as I doubt most of the people running LLM spam bots would be willing to pay $5k/month. I hope you are able to get an exception for a valid product use case.

@BGoat @replysocial Cost prohibitive at early-stage (likely *starts* at around $50k/mo)

Unfortunately this means the end of @replysocial. We used the "POST /2/tweets" endpoint to allow social media managers and support teams to reply to people talking about their companies and this kills off a large part of functionality around that. We did absolutely ZERO automated calls. Not a single line of AI/LLM code in the entire platform and no way to mass/batch interact with any posts in any way. The value proposition was that we centralized all the interactions you were naturally having with your customers. And yes, obviously it still works in the situation where people @-mention your company. But many times customers only mention you by name (and not by handle). That's extremely common for non-tech products. Despite their recent step forward in rolling out pay per usage pricing, the reality is that X is still very hostile to developers. I don't envy their position in dealing with spam/bots. It's an endless game of whack-a-mole and they've simply got to make tradeoffs to keep their platform from becoming ground zero for the dead internet theory. I get it. But as much as I wanted to build interesting/impactful tools *for humans* on X, I just can't keep throwing time/effort/money at a platform that continues to be so unstable. In addition, I'll be shutting down @BotBlockAI. It still functions perfectly fine but I have no idea what other API endpoints X will decide to cutoff overnight so I'd rather not keep pouring resources in to it.

Programmatic replies to X posts using the X API are now blocked unless the user is mentioned or their post is quoted. This is a great change!

@theo @trq212 @bcherny How is there still not a clear answer to this question from such a big name in the space? The silence is deafening. It is a simple question that deserves a simple answer.

@trq212 @bcherny Would love clarity and a direct answer on this if possible x.com/theo/status/20…

Hey @trq212 @bcherny - would love some clarification on this. Seems to contradict previous statements If I wanted to build my own OSS agentic code GUI, can I use the Agent SDK and let users bring their CC subs?

@nicoalbanese10 Love this update. I would love it even more if you could take a snapshot without stopping the sandbox 🙏

Let's go! Huge update from previous 7 day limit

Major Claude Code policy clear up from Anthropic: "Using OAuth tokens obtained through Claude Free, Pro, or Max accounts in any other product, tool, or service — including the Agent SDK — is not permitted"

@steipete @openclaw People who think you "failed" 43 times are way too focused on "hitting it big" with one project. To them, laying the foundation for something bigger is a waste of time. That is a sad mindset to have.

The funniest take is that I "failed" 43 times when people look at my GitHub repos and projects. Uhmm... no? Most of these are part of @openclaw, I had to build an army to make it useful. github.com/steipete/

This is a pretty big deal. The previous 7-day snapshot expiration meant building additional processes to backup snapshots to longer-term storage. That was a big reason I started working with Fly.io Sprites instead of Vercel Sandboxes. Might be worth revisiting.

Sprites from @flydotio are very cool... when they work. Running into major problems with Sprites becoming unresponsive with no way to recover. Attempts to restore to a checkpoint fail. No way to get at the filesystem. I hope they fix these issues quickly! They are unusable ATM.