531 posts tagged with "AI agents"
AI agents and autonomous systems
Hard Multi-Tenancy for Agent Workloads: What Apple's Kata plus VirtualClusters Recipe Costs
Apple's KubeCon recipe pairs per-tenant virtual control planes with Kata microVMs so untrusted agent code never shares a kernel with production tenants. The bill: about a second of cold start and 30-100 MB per sandbox, plus one small control plane per tenant.
Crossplane Graduates: AI Needs APIs, Not UIs — and What That Means for Your Self-Hosted PaaS
Crossplane's CNCF graduation validates the control-plane-first thesis: AI agents need declarative, reconciled APIs, not dashboards. Upbound's Modelplane proves it across clouds — but a fleet you own needs depth, not breadth.
Your PaaS Thought One Developer Meant One Deploy Queue. Then the Agent Fleet Showed Up.
A third of Cursor's own merged PRs are opened by autonomous cloud agents, and GitHub merged 43.2M PRs a month in 2025. One developer's agent fleet can now exhaust a whole account's daily deploy quota before lunch — here is the burst math and the four pipeline fixes: build queues, preview caps, isolated preview data, and merge triage.
MCP Has 10,000 Servers but Three Production Gaps: Identity Propagation, Timeout Budgets, and Structured Errors for Deploy Agents
Ten thousand MCP servers and 97 million monthly SDK downloads made connecting agents to tools trivial — but an agent that can deploy and roll back production needs scoped identity, timeout budgets, and machine-readable errors first. This post designs that deploy surface, tool by tool.
One GPU Per Sandbox: What Modal's Whole-Card Bet Costs Against E2B and Daytona's CPU-First Lanes
Modal puts whole GPUs inside sandboxes while E2B stays CPU-only and Daytona's GPUs are ephemeral-only. The reason is physical — Firecracker and gVisor cannot pass a GPU through — and the billing consequence is an 8x cost swing on GPU busy fraction alone.
Render Taught an AI Agent to Migrate Your Heroku App: What It Gets Right, What Still Needs You, and the Bill at the End
Render's open-source migration skill lets an AI agent move a Heroku app — dynos, config, Postgres — over MCP in minutes. A line-by-line test of the before/after bill, what the agent nails, where a human is still required, and the five capabilities your own fleet needs before migration becomes a prompt.
Your 8B Model Isn't Dumb. Its Harness Is: Forge's 53%-to-99% Guardrail Lesson for Platform Ops
Forge's May 2026 result took an 8B model from 53% to 99.3% on agentic tasks with guardrails alone — beating unguardrailed Claude Sonnet outright. What the retry-nudge math, the 75-point serving-backend swing, and a worked token-vs-hardware breakeven mean for running platform-ops agents on a self-hosted GPU.
Your Coding Agent Needs a Sandbox, Not a StatefulSet: gVisor vs Kata Under Kubernetes' New Sandbox Controller
Kubernetes SIG Apps shipped an upstream Sandbox controller for running untrusted agent code. A practical decision guide: what it replaces, whether gVisor or Kata fits your tenants, and what warm pools cost in idle capacity.
10,000 MCP Servers and No Human Wiring Them Up: What Server Cards and Stateless Operation Mean for Agents That Deploy Your Apps
MCP passed 10,000 public servers and is going stateless and self-describing, so agents will soon reach your deploy tools with no human wiring anything up. Here is the three-tier tool surface — observe freely, gate mutation, deny destruction — that survives contact with autonomous callers.