531 posts tagged with "AI agents"
AI agents and autonomous systems
Your Deploy-from-Chat MCP Server Has Too Many Tools: What Anthropic's 98.7% Token Cut Means for Agent-Operated Infrastructure
Anthropic cut a multi-step agent task from 150,000 tokens to 2,000 by having agents write code against MCP tools instead of calling them one by one. Here is what that 98.7% saving means for deploy-from-chat servers, and why every mutation still needs its own approval and audit entry.
One Gateway or Fifty Servers? What AWS and Google's Opposite MCP Bets Mean for Your Deploy Surface
AWS ships one MCP server with a fixed tool set while Google ships more than fifty, one per service. This post compares both designs on context cost, auth, versioning, and audit, then recommends a concrete tool vocabulary for a deploy surface agents can operate.
Claude Cowork Broke Containment Twice in July: VM Root on Windows, Host Files on Mac
In July 2026, Claude Cowork broke containment twice — a root chain on Windows and a kernel-to-host escape on Mac. This teardown walks both attack chains and maps the isolation tiers and hardening checklist your own agent sandboxes need.
Daytona's Core Went Private: What a Frozen Sandbox Repo Teaches About Who Owns Your Agent Isolation Layer
Daytona froze its open-source sandbox repo in June 2026, four months after a $24M Series A. What the freeze costs teams pinned to v0.190.0, how MinIO CE ran the same playbook, and a priced look at staying, renting, or owning your agent isolation layer.
Block the Training Bots, Welcome the Answer Bots: A robots.txt Default for the AI-Crawler Era
Publishers now block AI training bots while welcoming the same vendors' answering crawlers. A copy-paste three-tier robots.txt default for docs sites and tenant apps, with the traffic data and gotchas behind it.
The MCPA Is Here: What a 120-Minute, 5-Domain MCP Exam Says About Production MCP Operations
The Agentic AI Foundation's new MCPA exam puts half its weight on tool execution and security. A read of what that syllabus reveals about where MCP breaks in production, plus a five-domain checklist for hardening your own MCP server.
MCPwn: How nginx-ui's Unauthenticated MCP Endpoint Handed Attackers Full nginx Takeover — and the Checklist That Would Have Stopped It
nginx-ui shipped MCP support with one endpoint missing authentication, exposing 12 tools — including config writes with automatic reload — to anyone on the network. CVE-2026-33032 scored 9.8 and was exploited in the wild; here is how the takeover worked and a seven-item checklist for any panel adding agent access.
Node Lifecycle Conditions Land as Alpha in v1.37: The Machine-Readable Node Health Vocabulary Your Agent Operator Has Been Waiting For
Kubernetes v1.37 reserves five well-known Node conditions for drains, maintenance, and graceful shutdown. No core controller reads them yet — but for agent operators, a stable machine-readable vocabulary is the prerequisite, and self-hosted fleets can start publishing today.
Ownkube's Named AI Agents Run Your Ops — but the AWS Bill Is Still Yours
Ownkube pairs a Heroku-style deploy UX with named AI agents for cost, incident, scaling, and security — but its Production shape still runs on EKS. A worked 8-vCPU comparison puts the AWS route at $460–510 a month against €60–90 on owned hardware.