531 posts tagged with "AI agents"
AI agents and autonomous systems
Modal's $355M, Daytona's $24M, E2B's Billionth Sandbox: Where Agent-Infra Money Went in 2026
Modal raised $355M, Daytona raised $24M, and E2B passed a billion sandboxes — after the hyperscalers shipped their own. The scorecard, the per-sandbox-second math, and what teams running their own machines should rent, watch, or build.
The $6,500 AWS Agent Bill: Why Agents With Deploy Keys Need Guardrails, Not Bigger Budgets
An AI agent told to index a hobbyist network spun up five AWS instances and handed its operator a $6,531.30 bill. The wider 2026 incident ledger, why dashboards and bigger budgets keep failing, and the seven deploy-time guardrails — plus a cost dimension agents can read before they act — that would have stopped it.
Dapr Agents Hit 1.0: Idle as Actors, Execute in Sandboxes
Dapr Agents 1.0 models AI agents as virtual actors that shrink to kilobytes of state when idle and reactivate in milliseconds. A three-way cost table for 1,000 agents shows when actors beat always-on containers and checkpointed sandboxes — and where sandboxes still win.
Great at Code, Bad at the Terminal: Why Deploy Agents Need Structured APIs, Not Shell Access
Top coding agents score 74-79% on SWE-bench but drop twenty-plus points on Terminal-Bench's real shell tasks. Those failure modes map directly onto deploy incidents — and argue for scoped MCP tools over raw kubectl access.
Your Deploy CLI Should Behave Like kubectl: Kubeconfig Merges, Impersonation, and Agent-Operated Fleets
kubectl settled how fleet tools find their cluster and identity years ago: kubeconfig chains, first-wins merges, and impersonation as act-as-tenant. What your deploy CLI and agent tooling should borrow whole — and where kubectl parity stops.
E2B Gives Your Agent 1 Hour, Then Kills Its Sandbox: What Ephemeral-by-Design Costs a Long-Running Agent
E2B caps Hobby sandboxes at 1 hour, Pro at 24 hours, and auto-deletes paused sandboxes after 30 days. A day-by-day walkthrough shows what those clocks do to an agent running a multi-day task — and which persistent alternatives fit once disposability stops working.
Your Agent Wants Deploy Access. GitHub Just Showed What to Ask For First.
GitHub's Copilot app went GA with per-write permission prompts as the default and autopilot as the opt-in. What that validates about gating agent deploy, rollback, and scale calls — plus the 4-rung ladder and earned-autopilot path.
The GPU Isolation Fault Line: gVisor, Kata, and Edera Face Off Over Your Sandbox Pool
gVisor proxies CUDA calls, Kata hands the whole GPU to a VM, and Edera isolates tenants behind a hypervisor while sharing the card — what each model costs a self-hosted sandbox fleet in blast radius versus throughput, and which one to default to.
MCP Crosses 97 Million Installs: Who Actually Operates the Servers Behind Your Agent's Tool Calls
MCP's 97 million installs count client SDK downloads, not servers anyone operates. A custody accounting of vendor-hosted endpoints versus self-operated servers — token custody, poisoning audits, and what running your own MCP server takes.