Akshat Bubna

Akshat Bubna

Co-Founder & CTO at Modal

agentsinfrastructurestartupmodal

About Akshat Bubna

Akshat Bubna is co-founder and CTO of Modal, a compute platform built from scratch for AI workloads, with its own file systems, scheduler and container runtime. Modal runs inference, training, batch processing and sandboxes for untrusted code, and its sandboxes have become the runtime behind coding agents such as Ramp's internal Inspect agent and accounting agents such as Basis.

In an AI Engineer talk, he lays out the primitives agents need: a workspace with repo, data and services; persistence that keeps state indefinitely while running compute only when the agent works; elastic resources sized for variance rather than peak; GPUs on demand; shared filesystems for collaboration and memory; strict egress and credential control; and observability the agent itself can read.

Notable Positions

On Separating Brains From Hands

The harness, the loop that calls the model and runs tools, should live apart from the sandbox where tool calls execute. Harnesses are good at spinning up environments on demand, snapshotting them and returning later, which makes the pets-to-cattle analogy obsolete. See Agent Harness and Agent Sandbox.

On Persistence

Using Basis as the example: documents arrive, the agent works, then waits days for an accountant. State should persist, potentially indefinitely, while compute stops, through root filesystem snapshots, directory snapshots or shared volumes.

On Not Vibe Coding Infrastructure

Reliability under tail failures, continuous defense of the security boundary and scale (20 million snapshots a week, 10,000 sandbox starts per second for one customer) are reasons not to let an agent build your sandbox platform.

Key Quotes

  • "Every cell is backed by a different agent." (on Hebbia's spreadsheet product)
  • "You want to persist the state over time, potentially indefinitely. But you don't want to run the compute the entire time because compute is expensive." (on agent persistence)
  • "Great primitives let your agents do more." (closing thesis)

Related People