One of the most frustrating things about working with AI coding assistants has been their goldfish memory. You fix a subtle bug on Monday, explain to the agent exactly why the naive implementation doesn’t work, and by Tuesday you’re explaining it again from scratch. Session starts, context resets, you start over. Anthropic just shipped a fix for that. Claude Code now has auto-memory: it automatically creates and maintains a MEMORY.md file in your project, tracking debugging patterns, project-specific context, and your preferred working methods — across sessions, without any manual setup required. ...

Google doesn’t always announce its most important moves loudly. The rollout of an “agent step” update to Opal — Google Labs’ no-code visual agent builder — didn’t get a splashy keynote. But for anyone building enterprise AI agents in 2026, it quietly sets out a reference architecture worth studying carefully. Opal’s new agent step is now available to all users. And what it ships isn’t just a feature — it’s a working implementation of the design principles that serious enterprise agent builders have been converging on for the past 18 months. ...

If you’ve been running multi-agent AI systems and assuming your safety evaluations have you covered, a new study from five of the top research universities in the United States suggests you may be dangerously wrong. The paper, Agents of Chaos (arXiv:2602.20021), was produced by researchers from Stanford, Northwestern, Harvard, Carnegie Mellon, and Northeastern. Its core finding is stark: when autonomous AI agents interact peer-to-peer, individual failures don’t stay individual. They compound — triggering denial-of-service cascades, destroying servers, and consuming runaway resources in ways that single-agent safety evaluations simply cannot anticipate. ...

Anthropic just made Claude Cowork significantly more useful for anyone who wants their AI to keep working after they close the laptop. The desktop app now supports scheduled recurring tasks — meaning you can set Claude to run daily research, organize folders, compile summaries, or handle any repeatable workflow automatically, whether you’re at your desk or not. This is a meaningful shift. Cowork launched as a capable task runner, but it required you to be present to kick things off. With scheduled tasks, it starts to look a lot more like the always-on agent model that tools like OpenClaw have been built around from the start. ...

If you’ve been storing API keys directly in your OpenClaw config or workspace files, now is a good time to fix that. OpenClaw v2026.2.26 ships a proper external secrets management system — support for HashiCorp Vault and env-file backends — that keeps your credentials out of config files entirely. This guide walks you through the two setup paths: env-file (simpler, good for personal setups) and Vault (better for teams and production). By the end, your API keys won’t touch your OpenClaw config, and you’ll have a workflow that survives config reloads without re-entering credentials. ...

Getting OpenClaw running has always required a server — a Linux box, a VPS, a Raspberry Pi, something you configure, maintain, and keep online. MiniMax just changed that. MaxClaw, launched yesterday at maxclaw.ai, offers one-click cloud deployment of a fully functional OpenClaw instance in under 10 seconds. No server. No Docker. No config files. MiniMax is a Chinese AI unicorn that’s been building foundation models and infrastructure quietly while Anthropic and OpenAI dominate Western headlines. Their M2.5 model — a 229-billion parameter mixture-of-experts architecture — powers MaxClaw under the hood. But the product is OpenClaw-compatible: your agents, your skills, your integrations, running in their cloud. ...

If you run production AI agents, today is a good day to update. OpenClaw v2026.2.26 dropped overnight with a set of fixes that address some of the most frustrating pain points in long-running agentic pipelines — cron jobs dying silently, secrets leaking into environment dumps, and agents losing their memory between sessions. This isn’t a feature-focused release. It’s the kind of hardening update that makes everything else work better, and with 47 contributors and 11 discrete security hardening patches, it’s clearly been a community priority for a while. ...

Apple just crossed a line no major OS vendor had crossed before. With today’s release of Xcode 26.3, the company has shipped native, first-class support for autonomous AI coding agents — plugging Anthropic’s Claude and OpenAI’s Codex directly into the IDE via the Model Context Protocol (MCP). This is not a copilot autocomplete feature. These are full agents. What Xcode 26.3 Actually Does The new agent integration lets Claude and Codex do the kind of work a junior developer might spend a full morning on — autonomously. Here’s what agents can do inside Xcode 26.3: ...

Oasis Security disclosed a critical vulnerability chain in OpenClaw today that can enable full workstation compromise — initiated from a browser tab. SecurityScorecard found more than 40,000 OpenClaw gateways exposed to the public internet. If you’re running OpenClaw, this guide walks you through auditing your exposure and locking it down while you wait for an official patch. This is not a theoretical threat. Act now. Disclaimer: This guide reflects best practices as of 2026-02-26, based on the publicly available Oasis Security threat research. OpenClaw’s security team has acknowledged the report. Apply any official patches immediately when released, as they may supersede or extend these mitigations. ...

On the same day that Oasis Security disclosed a critical vulnerability chain in OpenClaw, and an MIT study found that most agentic AI systems have no documented shutdown controls, a credible new open-source project arrived that addresses both problems at the design level. IronCurtain — published today by Niels Provos, a security researcher with serious credentials (he’s known for work on OpenSSH and honeypot research) — is a model-independent security wrapper for LLM agents that enforces behavioral constraints without requiring changes to the underlying model. ...