Salesforce and Anthropic used Dreamforce 2026 to push MCP toward becoming the enterprise standard interface for agents. That shift creates a real infrastructure problem most platforms aren't built for.
Frontier models handle reasoning. But running continuous agentic work, the actual multi-step loops where an agent calls tools, checks results, and calls more tools, needs infrastructure that standard serverless and monolithic VMs were never designed for. Those were built for stateless web requests, not bursty, stateful, multi-agent loops executing tool calls in real time.
Here's how we built Kilawatt Cloud to be that hosting layer.
Three things an agent infrastructure layer actually needs
Stateful context holding: persistent bidirectional connections (Stdio or SSE) so a multi-step loop doesn't drop mid-task.
Low-latency burst compute: tight feedback cycles for code execution and database queries, without latency compounding across a pipeline.
Programmatic machine settlement: machine-to-machine micro-billing (HTTP 402 / X402) so an agent can pay for compute on its own, no human in the loop.
How the MCP pipeline works
An Express service boots paired with the MCP SDK. It registers tools like execute_sandboxed_compute across hardware tiers (CPU-high-freq, GPU-V100, GPU-A100). Incoming tool calls dispatch to that hardware and return structured JSON. SSE endpoints (/sse and /messages) keep the channel alive so long-running agent loops don't time out.
Payment is the auth
An agent calls the API. We respond with an HTTP 402 challenge, priced live off the real-time cost of whichever provider will actually serve the job, not a static rate card. The agent signs and sends a payment token. We verify it instantly, execute, and return results in milliseconds.
Real numbers, not theoretical ones
Fastest logged provisioning time across three GPU providers: 4.7 seconds.
Across 18 real, autonomous agent purchases, 16 (89%) needed an automatic failover when the first-choice provider couldn't deliver. Every one was caught and completed on a backup. Zero dropped requests.
Pricing is live, not fixed. Every job charges real cost plus margin at the moment of the request, so identical job specs can land at different prices depending on real market conditions.
What this means if you're building in this space
If you're building agent tooling, standardize on MCP instead of custom connectors. If you're building infrastructure, the leverage is in the stateful hosting layer and machine payments, not the model itself. If you're a technical leader, plan for high-concurrency, short-duration compute spikes, since that's what agent traffic actually looks like.
kilawattcloud.dev
github.com/kilawatt-mcp-server
Top comments (0)