🎯 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 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
| Tier | Where it runs | Role |
|---|---|---|
| Out-loop orchestrator | Your machine | Kicks off the work, then goes quiet. Top-level Claude Code (Opus). |
| In-sandbox orchestrator | Each VM | Runs the AI developer workflow (ADW) inside the box. |
| Software factory (ADW agents) | Inside the sandbox | The actual plan→build→test→review→document agents (via Pi). |
🔐 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
| Advantage | What it gives you |
|---|---|
| Isolation | Agents can't blow up your production system — strip their ability to touch AWS/GCP |
| Scale | One orchestrator spins up 5 sandboxes → 5 factories. "What if you could do this 100 times?" |
| Autonomy | Every agent owns a full computer, not a corner of yours. Blast radius is zero. |
🎲 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
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
| Tier | Definition | Examples |
|---|---|---|
| State-of-the-art | Top of the intelligence index (~56+) | Claude Opus 5, frontier models |
| Workhorse | Powerful, not quite SOTA, best economics (~50+) | DeepSeek V4 Flash, Gemini 3.6 Flash, GPT-5.6 Luna |
| Lightweight | The floor — runs on older GPU or Mac unified memory | On-device models |
💥 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
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 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
- 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.
- 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.
- 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.
- 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.
- Model stack, not model. Combine compute, don't select compute. DeepSeek V4 Flash delivers stock-market-crashing economics in the workhorse tier.
- 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."
- 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
- 📺 Original video — IndyDevDan's full build
- 💻 Factory In A Box codebase — open source
- 🖥️ exe.dev — agent sandboxes