An AI agent that writes code also wants to run it, and every run is untrusted software asking for a CPU, a filesystem, and a network. This full-time opening, one of a small number of remote agent infrastructure engineer jobs open to candidates worldwide, pays $125,000 per year (USD) to build the layer that lets thousands of those runs happen safely at once. It is fully remote and suits an infrastructure, platform, or backend engineer with about four years behind them who thinks about kernels, cgroups, and failure modes before features.
Why the project team needs a sandbox layer under its agents
The project team runs AI agents that open shells, install packages, and execute generated code. Each session must start in well under a second, stay walled off from every other session, and leave nothing behind. A shared-kernel container is too thin a boundary for code nobody has reviewed, so the platform runs Firecracker microVMs on Kubernetes, with long-running agent work coordinated by Temporal.
That stack needs an owner: someone who decides how many microVMs a node can carry, what happens when a workflow dies halfway through, and which egress rule stops a misbehaving agent from scanning the internal network.
Weekly work on the sandbox, workflow, and isolation stack
- Build the runtime that boots, snapshots, and tears down Firecracker microVMs for agent code execution.
- Run the Kubernetes node pools that host them, including scheduling, autoscaling, and capacity planning.
- Design durable Temporal workflows for agent sessions, with sensible retries, timeouts, and heartbeats.
- Harden isolation with seccomp profiles, cgroup limits, the Firecracker jailer and strict network policies.
- Instrument cold-start time, density and error rates per node and per tenant.
- Lead incident response and write blameless post-incident reviews that end in tested changes.
- Review designs and code in Go, Rust or Python.
The bar: four years, deep Kubernetes and a systems language
In plain terms, the agent infrastructure engineer job description asks you to bring:
- A bachelor's in computer science, engineering, or a similar field.
- Four or more years in infrastructure, platform, or backend engineering, running in production.
- Deep Kubernetes experience: scheduling, networking, and debugging nodes under pressure.
- Working familiarity with Temporal workflows and Firecracker microVMs.
- Strong Go, Rust, or Python.
- A solid grasp of Linux isolation and security: namespaces, cgroups, seccomp, and KVM.
Helpful, though not required
- Multi-tenant compute where tenants do not trust each other.
- Prometheus, Grafana, or OpenTelemetry in production.
Pay of $125,000 a year and the on-call rota
This role carries an agent infrastructure engineer salary of $125,000 USD per year, about ₹1.19 crore before tax at roughly ₹95 per dollar, a conversion that drifts with currency markets. As with most agent infrastructure engineer jobs remote worldwide, overlap hours and the on-call pattern are agreed with the hiring team during interviews. Bring a dependable laptop and connection, plus somewhere quiet for incident calls.
Design review, part one: workload and isolation for 4,000 agent sessions
The numbers below are illustrative.
Workload. At peak, about 4,000 code-execution sessions run at once. A typical session lasts six minutes, averages half a vCPU and spikes to two during package installs. Most need 512 MiB; one in twenty needs 2 GiB.
Isolation. Each session gets its own Firecracker microVM and guest kernel, started through the jailer so the VMM process sits in its own namespaces, cgroup, and seccomp filter. Egress is denied by default, and a per-tenant allowlist opens package mirrors only. Firecracker needs KVM, so the node pool is bare metal.
A reviewer comment from this stage: "Show me what an agent sees if it port-scans 10.0.0.0/8. I want the dropped packets on a dashboard, not a promise."
Design review, part two: scheduling, failure and the noisy-neighbor page
Scheduling. A daemon on each node keeps a warm pool of snapshotted microVMs, bringing median cold start to about 150 ms versus over a second for a full guest boot. With memory overcommit capped at 1.3 times physical RAM, a 64-core node carries about 110 sessions, so peak needs roughly 37 nodes plus headroom.
Failure. Each session is a Temporal workflow: create sandbox, run steps, collect artifacts, destroy sandbox. Destroy runs in a cleanup path even if the agent crashes, and a 30-second heartbeat timeout moves a lost worker's tasks elsewhere without leaking VMs.
Then there is observability. One afternoon, p99 cold start on a single node jumped from 210 ms to 1.4 seconds. Per-VM disk metrics showed one agent compiling a large native library and saturating the NVMe queue. The fix was a token-bucket rate limit on each microVM's block device, plus an alert on per-VM I/O wait. The most useful line in the eleven-line post-incident note said the limit had been in the design document and never deployed.
How backend engineers move into agent infrastructure
Most people here were backend or platform engineers who kept volunteering for cluster work. If you wonder how to become agent infrastructure engineer material from there, build a home lab: a bare-metal box or a cloud instance with nested virtualization, Firecracker booting a minimal guest, a small Kubernetes cluster scheduling it, and a Temporal worker driving the lifecycle. Break it on purpose and write up the fix.
No single agent infrastructure engineer course exists yet; the free Temporal 101 course on Temporal's learning site is a good start. For an agent infrastructure engineer certification, the closest fits are the Linux Foundation's CKA and CKS, and CKS maps directly to the isolation side of this job. A public repository with a working sandbox counts for more.
The core agent infrastructure engineer skills are Linux internals, Kubernetes, one systems language, and calm debugging. Entry-level remote agent infrastructure engineer jobs are rare, and openings for people with no experience hardly exist; a junior backend or SRE role comes first.
Interview questions in remote agent infrastructure engineer jobs, and resume proof
A strong agent infrastructure engineer resume leads with numbers from systems you ran: "Cut p99 pod start from 9 s to 2 s on a 300-node cluster" beats a list of tools. Typical agent infrastructure engineer interview questions:
- Why choose a microVM over gVisor or a plain container for untrusted agent code, and what do you give up?
- A Temporal activity times out after the sandbox was created but before its ID was recorded. How do you avoid a leaked VM?
- Density per node has dropped 30% in a month. How do you find out why?
- Write a NetworkPolicy that lets a sandbox reach a package mirror and nothing else.
Agent infrastructure versus platform engineering, and where demand sits
A platform engineer builds paved roads for the company's own developers. This role builds compute for code nobody at the company wrote or trusts, so threat model, latency budget, and cost per session drive every decision. The skills overlap heavily. Demand has grown as coding agents reached real products, and companies such as E2B and Modal now sell agent sandboxes as a service, though openings still arrive in uneven waves.
Fake infrastructure offers, and applying for this one free
It is a scam if a recruiter wants money up front for registration or a "training kit", asks you to read out a bank OTP or approve a UPI request, keeps every conversation inside Telegram or WhatsApp, or offers a senior engineering post with no technical interview. Nothing about this listing costs you money. Be wary of agent infrastructure engineer work-from-home jobs elsewhere with vague duties and no named stack.
To apply, send a resume that highlights your Kubernetes, isolation, and on-call experience, along with any sandbox work you can share. Applications are free to submit; shortlisted candidates move to technical interviews, and selection is not guaranteed. Among remote agent infrastructure engineer jobs hiring now, trust the ones that name their stack and test you on it.
