blog
Blog — page 8 of 39
Posts 169–192 of 928, newest first.
The anatomy of an AI agent's bill: tokens vs compute vs egress
Three invoices, no shared key, and the biggest line grows with the square of your turn count. Here is one agent run taken apart, with a lever for each cost line.
Ajay KumarHow to give a Strands agent a code execution tool
Strands ships a sandbox interface and names its default NotASandboxLocalEnvironment. Here is how to implement that interface against a real microVM, and every decision that follows.
Ajay KumarTekton Steps, MicroVM Bodies: Isolating Untrusted Tasks
A TaskRun is a Pod. That is Tekton's best feature and its sharpest constraint — because everything inside a Pod is inside one trust boundary, including the step running a stranger's build script.
Ajay KumarCross-Compilation Build Farms: One MicroVM Per Target
One binary becomes a matrix the moment somebody asks for arm64. Emulation lies about the parts you care about, cross-compiling breaks on the sysroot — here is the third answer.
Ajay KumarOSS License Compliance Scanning, One MicroVM Per Scan
You cannot produce a defensible license report from a lockfile — the evidence lives inside the tarballs. Which means something of yours has to expand a few thousand strangers' archives.
Ajay KumarThe Best Prefect Hosting Platforms in 2026
Almost every 'how do I host Prefect' thread is two questions wearing one coat: where does the server live, and where do flow runs actually execute? They have different answers, and the second one is the one that costs money.
Ajay KumarThe Best Metabase Hosting Platforms in 2026
Running metabase.jar takes five seconds. Running it in production means an application database that is not H2, a heap the kernel agrees with, and a SQL console you have not accidentally published.
Ajay KumarThe best ways to host MLflow in 2026
MLflow looks like a web app and is actually a database, a bucket, and a code-distribution channel wearing one name. An honest look at where to put each part — including where we're the wrong answer.
Ajay KumarConntrack Table Limits: The Shared Resource Under Your MicroVM Fleet
Your guests have separate kernels, separate namespaces and separate subnets. They share one fixed-size hash table, and one port scanner can fill it for everybody.
Ajay KumarHyper-Threading and microVM Isolation: the SMT Decision
The most expensive security decision in multi-tenant compute is one nobody puts in a slide: do you leave SMT on? Two threads on one core share the L1, the TLBs, the branch predictors and the execution ports — below the level where a hypervisor can isolate anything. Here's the topology, the sysfs evidence, and the four rungs of the decision.
Ajay KumarHow to deploy a Qwik app
Qwik's adapter is not a detail you pick at the end — it is the deployment model itself. Here is what resumability changes about the server, and how to ship the Node path.
Ajay KumarHow to Self-Host authentik for SSO
authentik is a genuinely good open-source identity provider with an unusual abstraction. Here is the component model, the concepts that trip people up, and the operational promises you are making.
Ajay KumarWhat Is an AI Agent?
An AI agent is a language model in a loop: it proposes an action, your code runs it, the result goes back into context, repeat. The loop is twenty lines. The environment those actions run in is the actual product.
Ajay KumarEphemeral Databases for AI Agents
Give an agent a run_sql tool and it will use it — including the DELETE it emits while debugging its own step. The fix isn't a better statement filter; it's a database that exists for one task and is deleted when the task ends.
Ajay KumarThe best Inngest alternatives in 2026
Most teams shopping for an Inngest alternative do not want a different step DSL. They want the same work expressed as an ordinary long-running process they own.
Ajay KumarThe Best Dagster Hosting Platforms in 2026
You do not deploy 'Dagster'. You deploy a webserver, a daemon that must be a singleton, one or more code-location servers, and a Postgres you cannot skip. Here is how the hosting options handle that, honestly.
Ajay KumarThe Best Ways to Host Grafana Loki in 2026
Loki's whole pitch is that logs go in object storage and only labels get indexed. That is what makes it cheap, and it is also the exact thing people break in week one.
Ajay KumarThe best Typesense hosting platforms in 2026
Typesense keeps its index in memory. Every hosting decision below — node size, cluster shape, monthly bill, and whether you should cluster at all — falls out of that one sentence.
Ajay Kumarffmpeg Command Injection: The Filename Is Input Too
A filename is user input. If you build your ffmpeg command with an f-string, so is your shell — and fixing that still leaves a C parser eating attacker bytes.
Ajay KumarBuildkite Agents on MicroVMs: One Job, One Machine
Buildkite's whole premise is that you supply the compute. That makes 'what do agents run on?' a product decision, not a detail — and containers answer it badly for CI.
Ajay KumarRunning Nix Builds Inside a Disposable VM
Nix's sandbox is a hermeticity fence, not a hypervisor — a hostile derivation still runs on your kernel. Here's how to put Nix inside a disposable VM without a cold /nix/store making every build miserable.
Ajay KumarWebhook Replay and Relay: Ingest First, Process Later
The provider retried for eleven hours, gave up, and the event is gone with no record you can inspect. The fix is boring and it is one afternoon of work: ingest first, process later.
Ajay KumarBuilding Container Images for Untrusted Repos Without Privileged Docker
Every RUN line in a user's Dockerfile is a shell command on your builder, as root, at build time. The build is the untrusted workload — not the thing you build.
Ajay KumarBuildKit vs Kaniko vs microVMs for Untrusted Image Builds
Every image builder is a code-execution engine wearing a build tool's clothes. Here is the honest comparison of the three ways teams solve that.
Ajay Kumar