Factory's FDE lead: enterprise agents should run as their own apps inside the customer's VPC, not borrow a user's identity
Nikhil Harithas, who leads forward deployed engineering at Factory, told AI Engineer that big customers like Blackstone keep asking who an agent acts for and where it runs, and demoed Open Compute, an egress-only control layer for running Factory's Droid agent on customer infrastructure.
Nikhil Harithas, head of forward deployed engineering at AI coding-agent company Factory, argued in an AI Engineer session (written up by StartupHub.ai on Nov 3) that agents should get their own service identities instead of piggybacking on a human user's permissions. He said two questions come up with every large customer: who the agent is acting on behalf of, and where it actually runs.
His team's work with Blackstone, which wanted its own compute, inference, identity and single tenancy, pushed Factory to build Open Compute. A controller inside the customer's VPC uses drivers for EC2, EKS, Kubernetes or sandbox providers to start microVMs that dial out to a relay, so no inbound connection is needed. It also handles pause, resume, snapshots and forking, which matters because Factory missions can outlast AWS microVMs' eight-hour cap. Factory says it plans to open source the layer.
Why it matters for FDEs: this is a concrete look at what deployment engineers at agent companies spend their time on. Identity, network ingress, secrets and compute placement inside regulated customers are the blockers, and the FDE team ends up turning those field lessons into product.
Source: www.startuphub.ai