Sandboxes are dead — the agent needs its own datacenter

The sandbox's obituary

E2B's argument is blunt: the sandbox, the unit of agent compute for three years, no longer fits the work. The company maps three waves — copilots for code completion, then agents that built full web applications, and now a third wave of agents controlling entire computers for long-running tasks. What those agents need, E2B says, is not a bigger sandbox but an 'agent-driven private cloud': infrastructure with dedicated storage, firewalls, and networking between sandboxes that adapts to each agent's resource and tool requirements. The claim is backed by scale — E2B says it has processed billions of agent runs and generated tens of billions of traces, enough to predict an agent's tool usage and resource consumption before execution. The sandbox was built for code snippets. The agent now wants a machine, a network, and a future. m:…7407

Bring your own everything

Replit is meeting the enterprise version of the same demand. Its new Bring Your Own Cloud offering lets enterprise teams keep development and production environments inside their own cloud tenants, with three deployment models: Full BYOC, currently in private preview, brings the entire Replit platform and published project hosting into the customer's tenant; BYOC for Published App, in early access, hosts only published projects there; and Image Exports, also in early access, let teams self-host Replit-built apps with Docker images, Terraform files, and SBOMs. The BYOC options support Microsoft Azure, Cloud, and Amazon Web Services, with mTLS-encrypted tunnels and customer-managed encryption keys for full physical isolation. Most telling is the Image Export detail: the exported app does not need to communicate back to the Replit platform at all. The platform is learning to sell itself as software you run, not a service you visit — because enterprise agents cannot live on someone else's cloud. m:…7523

The filesystem learns to run

Modal Labs shows what the substrate work looks like at the systems level. Its VM Sandboxes, built for what it calls sophisticated agentic workloads at scale, were being choked by the filesystem: under the old virtio-fs stack, a recursive chown took over a minute versus four seconds natively, with every operation requiring round trips from guest VM to host. The new ImageFS-blk moves to a block-device approach that keeps metadata operations inside the guest kernel — chown dropped from 9.87 seconds to 0.21, and a bun install benchmark fell from 12.5 seconds to 1.07, while containers still cold-start in under a second. When agents install packages, traverse trees, and hammer the filesystem thousands of times per run, the filesystem is not a detail. It is the product. m:…6969

The bill comes due

Pydantic's September is the cautionary tale that makes the buildout legible. Its CI bill was $3,990 in June, $10,778 in July, and $18,200 in August — driven, the company says, by AI-driven development: the Logfire team merged 2,032 pull requests in August 2026, versus 365 in the same 30 days in 2025 for a same-sized team. Runner minutes kept rising even after the fix, from 1.23M to 1.62M between August and September. Moving CI to RunsOn on September 1 — ephemeral EC2 runners in Pydantic's own AWS account, Graviton for most jobs, spot where retries are cheap — brought the September bill down to $5,700. This suggests the real story of agent infrastructure is cost control, not just capability: agents multiply work faster than they multiply budgets, and the companies that survive the transition will be the ones that re-engineer the substrate — the cloud tenancy, the filesystem, the runner fleet — around the agent as the unit of compute. m:…9429