Nikhil Harithas, who leads forward-deployed engineering at Factory, describes its agent infrastructure in this 17-minute AI Engineer talk.

Enterprise requirements break the defaults

  • Harithas uses Blackstone to explain requirements for customer-owned compute, inference, identity, and single-tenant deployment.
  • Delegating a Jira ticket to Droid, Factory’s coding agent, raises two questions: whose authority does it use, and where does it execute?
  • He argues for application identities and explicit permissions rather than indefinitely borrowing a developer’s identity.

Separate compute from the product’s control layer

  • Factory calls its proposed reusable lifecycle layer Open Compute; drivers adapt it to EC2, Kubernetes, or sandbox providers.
  • A compute gateway requests a machine through a controller in the customer’s VPC, an isolated cloud network.
  • The machine starts the agent daemon and opens an outbound WebSocket to Factory’s relay for client communication.
  • This avoids new inbound connections after setup while retaining an external control plane and a communication path to it.
  • Bring-your-own-key inference can use an approved provider arrangement such as Bedrock, within the customer’s chosen network setup.

Manage machines through their whole life

  • Installing an agent leaves someone responsible for starting, pausing, resuming, destroying, and locating its compute.
  • Templates and snapshots capture an environment’s build requirements; forks let another agent branch from that prepared state.
  • Harithas discusses customer-owned snapshot storage and eventual secret-store integration as further extension points.
  • Compute is the reusable unit; Factory’s conversational sessions stay a separate product concept.

Reliability lives below the agent loop

  • A cloud API can acknowledge a request without producing a usable machine; dropped acknowledgments create further uncertainty.
  • The platform must track machine identity, route users to the right sandbox, and reconcile retries.
  • Long-running missions may outlive a provider’s sandbox lifetime, requiring snapshotting and restoration onto replacement compute.
  • A driver specification gives coding agents a concrete interface to implement, though satisfying it alone does not prove operational reliability.

Caveat: Open Compute is described as a private-preview project planned for open sourcing. This architecture talk does not establish a released repository or prove complete data confinement.

“Who is an agent acting on behalf of and where does it run?” — Nikhil Harithas, 2:41