Software Factory in a Box

Engineers… Your Software Factory NEEDS Agent Sandboxes to SCALE (exe.dev)

🎬 IndyDevDan 📅 Aug 10, 2026 ⏱ 37:14
agent sandboxes software factory agentic engineering exe.dev best of N

🎯 The Core Idea — You Are the Bottleneck

"Your software factory can unlock an unprecedented level of engineering results. But there's a massive roadblock every agentic engineer runs into: where should your software factory run?" Most engineers allocate a tiny corner of their own computer for their agents, or lean too heavily on CI/CD. The best engineers do something else — they put the whole factory inside an agent sandbox. 0:00

The raw reality: "If you are inside the loop, you are the bottleneck." An agent sandbox gives three advantages a container can't: true isolation, insane scale, and agency. "When you use agent sandboxes, you're getting isolation, you're getting scale, and you're getting autonomy. Every single agent owns the computer. They're not just running on a little corner of yours, interfering with your work, you interfering with theirs." 1:09

The demo uses exe.dev as the sandbox tool — "Computers for developers and agents. Durable sandboxes, fast, secure, and shareable. The agent sandbox is a developer device for your agents." 4:12

🏗️ The Three-Tier Architecture

IndyDevDan builds a three-tier system. Understanding why there are two orchestrators is the key to the whole design: 13:12

TierWhere it runsRole
Out-loop orchestratorYour machineKicks off the work, then goes quiet. Top-level Claude Code (Opus).
In-sandbox orchestratorEach VMRuns the AI developer workflow (ADW) inside the box.
Software factory (ADW agents)Inside the sandboxThe actual plan→build→test→review→document agents (via Pi).
Why two orchestrators? "Our top-level orchestrator agent has stopped. It's not doing anything. All it did was kick off the orchestrators inside our sandbox. You get a little bit less control if you skip the in-box orchestrator — it shouldn't be doing anything but running and kicking off the workflows. The key is keeping it simple." The point of the whole thing: you show up at the beginning and the end — the planning and the reviewing — "and nowhere in between." 27:08

🔐 Isolation, Scale, Autonomy

"A container gets you isolation only. A sandbox gets you all three, and autonomy is the one that actually moves the needle." 33:18

AdvantageWhat it gives you
IsolationAgents can't blow up your production system — strip their ability to touch AWS/GCP
ScaleOne orchestrator spins up 5 sandboxes → 5 factories. "What if you could do this 100 times?"
AutonomyEvery agent owns a full computer, not a corner of yours. Blast radius is zero.
The blunt advice: "If you're using an agent to directly modify the application-layer code, you are wasting time. There are very few exceptions to that. I'm not trying to talk above you — I'm saying this so you understand where the leverage is." 35:00

🎲 Best of N — Five Futures, One Pick

Instead of running one agent once, IndyDevDan runs five different agent configurations — Default, Frontier, Deepest Seek (all DeepSeek V4 Flash), Open Weights (GLM 5.2, Kimi K3), and Top Speed (Gemini 3.6 Flash, DS V4 Flash, GPT-5.6 Luna) — each in its own sandbox, all solving the same problem against one prompt. 3:00

The pattern: "Say you have a great idea, or multiple directions. In our example we pass in one singular prompt — a redesign. Not just running one agent: a pipeline of agents plus code." One prompt fires the orchestrator, which spins up in-sandbox orchestrators to run the ADW system, and you pull the best result. "Best of N gives you the best of some set of results — and you configure your model selection. Every run is a different team of agents." 10:06

A side effect: as you run different variants of your factory, you harvest data — "how good is this configuration? Is it getting me results cheaper? Do I need a state-of-the-art system?" One config (Kimi K3, open weights) actually failed on a JSON format issue — "that's okay, another reason best-of-N is so important." 14:02

📊 Model Stack, Not Model

"If you're still fixated on models, you are behind. We're entering an age of abundant compute — now it's about what model plus what code do you need to do the job." The model stack has three tiers: 5:34

TierDefinitionExamples
State-of-the-artTop of the intelligence index (~56+)Claude Opus 5, frontier models
WorkhorsePowerful, not quite SOTA, best economics (~50+)DeepSeek V4 Flash, Gemini 3.6 Flash, GPT-5.6 Luna
LightweightThe floor — runs on older GPU or Mac unified memoryOn-device models
"Combine compute, don't select compute." The DeepSeek V4 Flash (0731) build — "Deepest Seek" — "is absurd for its price." It's in the workhorse A-tier, delivering "stock-market-crashing economics." "These fast, relatively cheap workhorse models can do a lot of the work you're probably doing right now." 16:50

💥 Bounded Blast Radius

"Disposable keys, bounded blast radius. Every sandbox gets a freshly minted OpenRouter provisioning key with a hard spend cap at mount, and that key gets revoked at teardown." IndyDevDan limits his to $50, then kills the key when the boxes tear down. 29:00

"The agent can do anything it wants. What it can't do is touch your AWS, your prod, or your wallet. The blast radius IS the box." These are ephemeral sandboxes — spin them up, do work, spin them down. Mount → run → teardown, two scripts. 29:10

And "out-loop does NOT mean blind": every sandbox exposes ports — a live URL for the running app plus a private URL for the factory observability view. "Watch the gate checks, the token spend, the agent harness config, the failures. You can SSH straight into any box mid-run and talk to the orchestrator that's already working. Measure it, or you can't improve it." 22:15

⚖️ Agentic Engineering vs Vibe Coding

"Be careful how much of your engineering work you're handing off to these AI labs. Don't outsource your thinking. Stay close to the engineering." 18:40

The fine line: "Vibe coding is not knowing what your system does and not looking. Agentic engineering is knowing what your system does so well you don't have to look. Big difference. A vibe coder and an agentic engineer might look like they're doing the same thing — but the agentic engineer is thinking in systems, thinking about the hundredth and thousandth run, thinking in observability, reusability, isolation, scale." 18:55

"The software factory is deterministic code wrapped around non-deterministic agents. That's what separates agentic engineering from vibe coding." The next leap will take work: "all the easy leverage has already been sucked up. Everyone has a terminal agent now. Almost nobody has a factory in a box." 36:30

✅ Key Takeaways

  1. If you're in the loop, you're the bottleneck. Great agentic engineering means showing up at the beginning and end — planning, reviewing, prompting, validating — and nowhere in between.
  2. A container gets you isolation only; a sandbox gets you isolation, scale, and autonomy. Autonomy is the one that moves the needle: every agent owns a full computer.
  3. The three-tier architecture is the unlock. Out-loop orchestrator (your machine) → in-sandbox orchestrator (each VM) → the software factory's ADW agents. Your top-level agent kicks off the work and goes quiet.
  4. Best of N — one prompt, five factories, five versions of the product. Pick the winner, move on, and harvest cost/quality data on every config for free.
  5. Model stack, not model. Combine compute, don't select compute. DeepSeek V4 Flash delivers stock-market-crashing economics in the workhorse tier.
  6. Bound the blast radius. Fresh OpenRouter provisioning keys with a spend cap at mount, revoked at teardown. "The agent can do anything. What it can't do is touch your AWS, your prod, or your wallet."
  7. Agentic engineering ≠ vibe coding. The software factory is deterministic code wrapped around non-deterministic agents — you know what the system does so well you don't have to look.

🔗 Resources & Links

📍 Timestamp Index

0:00 The bottleneck problem — where should your factory run?
1:09 Isolation, scale, autonomy — why not just a container?
3:00 Best of N — five agent configs, one prompt
4:12 exe.dev — "computers for developers and agents"
5:34 Model stack: SOTA / workhorse / lightweight tiers
10:06 Best of N pattern explained
13:12 Three-tier architecture — two orchestrators
18:40 Don't outsource your thinking
22:15 SSH into any box mid-run — out-loop ≠ blind
29:00 Disposable keys, bounded blast radius
33:18 Why sandboxes matter — isolation, scale, autonomy
36:30 The age of abundant compute — act like a function
☰ View all