Project
Microstax
An agent-native environment runtime — isolated, governed Kubernetes sandboxes for human developers and autonomous AI agents.
< 60s spin-up · 8+ hrs saved per developer per week · onboarding cut from 2 weeks to 1 day
What it does
Microstax is a control plane for on-demand developer environments. It provisions isolated Kubernetes namespaces from Blueprint definitions, automates the full lifecycle (create, share, debug, tear down), and integrates with CI/CD so every PR gets a branch-specific environment that inherits from a shared reference baseline.
The same runtime serves human developers and autonomous AI agents — both get governed sandboxes with NetworkPolicy enforcement, real-time logs, and auto-shutdown.
How to inspect it
- Product leaders: Look at the shared-reference / overlay workflow. The product decision was not "more customization"; it was faster time-to-first-success for a new developer.
- Engineering leaders: Inspect the lifecycle constraints: namespace isolation, NetworkPolicy, TTL cleanup, PR environments, and observability. Those are the difference between a demo environment and infrastructure your team can support.
- Operators: Ask what happens when a sandbox leaks resources, a PR environment fails to boot, or an agent needs a governed runtime. Those failure paths shaped the product more than the happy path did.
What we built, end to end
- The Kubernetes control plane (k3s, namespace orchestration, NetworkPolicy enforcement).
- The Blueprint DSL — Docker Compose compatible, with overlays for shared reference environments.
- The lifecycle automation: TTL-based cleanup, snapshots, sharable URLs.
- The integrations: GitHub Actions for PR-based environments, VS Code extension, REST API, Prometheus metrics.
- The first-class agent support: LangGraph + MCP Server primitives.
Discovery to first paying user: 14 weeks. Three senior engineers, full-time.
The hard call we made in week 3
We started Microstax with the assumption that every team wanted custom Blueprints from day one. The first pilot user told us in week 3 that they would not invest two days writing their own Blueprint just to try the product. We threw out the custom-first model and built the shared-reference / overlay system the night after that call. Half the team disagreed at the time. The number we now lead with — onboarding in 1 day instead of 2 weeks — would not exist without that pivot.
What we cut
- A custom in-browser IDE. Real users wanted to use their own VS Code; the in-browser IDE was a feature in search of a story. Cut at week 5.
- A multi-cloud abstraction layer. The first three customers all ran on GKE. We picked GKE-native primitives and got 6 weeks back.
- Real-time collaborative editing inside a shared sandbox. Useful, but not why anyone was paying. Filed under "v2."
Outcomes in production
- Spin-up under 60 seconds at p95.
- 8+ hours saved per developer per week (measured against their previous staging-slot flow).
- 100% team isolation — no more shared-staging contention.
- New engineer productive in 1 day, down from 2 weeks.
- ~70% reduction in infrastructure cost per environment with shared-overlay reuse.
These numbers come from the early customer workflow we replaced, not from a generic benchmark. They are useful because they describe the before-and-after operating model: fewer shared staging collisions, faster environment creation, and less repeated setup work.
Why this matters for your project
If you’re building developer-facing infrastructure — and especially if AI agents are going to be one of the "users" of that infrastructure — the design decisions we already made on Microstax are decisions you won’t have to re-litigate. We can show you the failure modes before you hit them.
If your project sits somewhere else entirely, the same discipline applies: write the eval first, kill the wrong assumption in week 3, ship to one real user by week 2.
Have a bounded AI system to build?
30 minutes. Problem, owner, budget, date, and success criteria.