blog
Blog — page 15 of 39
Posts 337–360 of 928, newest first.
How to Give a smolagents Agent a Code Execution Sandbox
In smolagents, every agent action is Python. That makes the code executor the single most important security decision in the whole framework — here is how to get it right.
Ajay KumarHow to Generate a Typed API Client From an OpenAPI Spec
Hand-written HTTP wrappers rot. Here is how to generate a typed client from an OpenAPI spec instead, and the three places generators reliably fall over.
Ajay KumarHow to Receive Webhooks for Deploys and Quota Events
Polling an API every thirty seconds to find out whether a deploy finished is a habit worth breaking. Here is the webhook version, including signature verification done right.
Ajay KumarHow to Expose a Sandbox Port on a Public URL
Your agent just started a dev server inside a VM. Getting a link to it is one line — but the URL is a capability, and that is worth understanding before you paste it into Slack.
Ajay KumarOne microVM per Terraform run: isolating IaC per tenant
A Terraform runner downloads third-party Go binaries and executes them with a customer's cloud credentials in its environment. If you run every tenant's HCL in one shared pod, you have built a credential blender. Here is the per-run microVM shape instead.
Ajay KumarYour Backup Is a Hypothesis: DR Drills in Disposable microVMs
Almost nobody tests their restore path, because testing it properly means restoring production data somewhere real — and the available venues are production itself and a staging environment that drifted two years ago. Here is a third one.
Ajay KumarOne microVM Per Sync Run: Isolating SCIM Connectors
To sell upmarket you have to sync users from every customer's Okta, Entra ID, Workspace and LDAP — which means holding hundreds of long-lived identity-provider credentials in one process. Here is what it looks like when each sync run gets its own microVM instead.
Ajay KumarYour Free Tier Is a Mining Pool: Stopping Cryptomining Abuse
The day you offer strangers free compute with a shell, you have opened a very generous mining pool. The miners will find it before your first paying customer does — here is what actually stops them.
Ajay Kumaruserfaultfd in Production: How the Pager Fails
With userfaultfd you have volunteered to be the pager for someone else's kernel. When you stop answering, nothing returns an error — you get a guest that is alive, scheduled, and permanently stopped mid-instruction.
Ajay KumarSparse Files and Hole Punching: Why Your Snapshot Lies About Its Size
A Firecracker memory snapshot for a 4 GiB guest is a 4 GiB file that may be occupying a small fraction of that on disk. That gap is one of the most useful facts about snapshot storage — and one of the easiest to destroy by accident with a single cp.
Ajay KumarMAP_PRIVATE vs MAP_SHARED: The mmap Flag That Decides Everything
One flag in one syscall decides whether your writes are invisible scratch, visible to every other process, or durable on disk. Here's what MAP_PRIVATE and MAP_SHARED actually do to a page, why VMMs, databases and caches pick opposite sides, and the traps — SIGBUS on truncate, msync that does nothing, and RSS numbers that lie.
Ajay KumarMount namespaces and pivot_root, explained for sandboxing: how filesystem isolation really works
A mount namespace gives a process its own copy of the mount table, and pivot_root replaces its root so there is nothing left to escape to. Both are excellent. Neither one gives the code a different kernel.
Ajay KumarHow to give a Google ADK agent a code execution tool
In ADK a tool is a Python function and its docstring is the prompt. The hard part isn't the wiring — it's deciding how long the sandbox lives and what you hand back to the model.
Ajay KumarBest Temporal Hosting Platforms in 2026
Everyone shops for a place to run the Temporal cluster. The part that actually decides your architecture is where the workers live — long-lived pollers that must not be killed mid-activity.
Ajay KumarEvery Listing Is a Stranger's Agent: Isolating a Marketplace
You launched an agent marketplace. Every listing is code a stranger wrote, driven by a model, calling tools you didn't write, against your customers' credentials. Here's the isolation shape that survives that.
Ajay KumarHostile and Precious: Per-Case microVMs for Forensics and E-Discovery
The PST you have to open is full of the malware that caused the incident, and you are also legally obligated not to change a single byte of it. One VM per matter is the shape that satisfies both.
Ajay KumarOne Demo per Prospect: Sales Environments You Can Throw Away
Two AEs demo at the same time and one of them watches their data change live on the call. The fix is structural: bake one golden demo — app plus seeded database — and fork it per prospect, so "give me a clean environment" is a button and not a ticket.
Ajay KumarRun the PoC: Vulnerability Triage in Disposable microVMs
A bug bounty report is a stranger's script plus a claim, and the only way to check the claim is to run the script — against a system you have deliberately made vulnerable. Here is the per-report microVM shape that makes that a routine Tuesday instead of an incident.
Ajay KumarYour Firecracker Snapshot Restore Failed: A Field Guide
A restore either works in tens of milliseconds or fails with an error that tells you almost nothing. Here are the seven classes of Firecracker snapshot restore failure, what each looks like, and how to bisect your way to the real cause.
Ajay KumarYou Ship the Kernel: Firecracker Guest 5.10 vs 6.1
In a microVM, the guest kernel is a build artifact of your platform, not something a distro picks for you. Here's the honest 5.10 vs 6.1 trade — and the snapshot re-bake nobody budgets for.
Ajay KumarShould You Compress Firecracker Memory Snapshots?
Compression looks like free money when your snapshot bucket is measured in terabytes. Then you discover that a compressed stream has no byte N — and your lazy 49ms restore turns into reading four gigabytes you were never going to touch.
Ajay KumarBest n8n Hosting Platforms in 2026: A Self-Hoster's Guide
You've decided to self-host n8n. Now you have to pick infrastructure for a stateful, always-on, long-running service that also happens to execute arbitrary JavaScript for a living.
Ajay KumarBest LangGraph Deployment Platforms in 2026
Your graph works in the notebook. Now it has to survive a deploy, a crash, and a human who takes three days to approve step 7. Here are the requirements that eliminate most platforms, and the seven that are left.
Ajay KumarFirecracker vs Proxmox: a VMM Is Not a Platform
Proxmox VE manages VMs. Firecracker runs one. The comparison is a layer mismatch — but the question underneath it is real, and the answer usually comes down to whether your guests are pets or cattle.
Ajay Kumar