Introduction

As AI agents become increasingly capable, companies are racing to integrate them into their workflows. But building a secure, scalable agentic infrastructure that actually works in production is difficult even for straightforward use cases.

‍

Faire, a leading wholesale marketplace, faced this challenge when they built an internal agent system that could write code, answer knowledge base queries, and execute arbitrary actions with access to internal tools.

‍

The project was complex and multifaceted. But with Crafting, Faire realized and surpassed this initial vision. The result? A production-ready agentic stack built on Crafting that powers their entire engineering organization and all their internal agents. The same infrastructure powering their engineers now powers their agents, validating code against real production dependencies, not just running it in isolation.

‍

The Challenge: Building Agents That Scale

Faire's engineering team had a clear goal: create an agentic system that could autonomously perform the full spectrum of software production. This meant supporting three core use cases:

• Code Generation: Agents that could write, review, and modify code across their entire codebase
• Knowledge Base Queries: Intelligent question-answering capacities that could pull from internal documentation, code repositories, and tribal knowledge
• Arbitrary Tool Use: Agents that could take action using any combination of internal tools, APIs, and cloud resources

‍

But supporting these use cases at scale introduced serious challenges:

• Security: How do you give agents access to internal systems and API keys without risking credential leaks?
• Scalability: How do you provision ephemeral execution environments that can handle unpredictable workloads?
• Flexibility: How do you support diverse agent personas, toolchains, and MCP servers without creating maintenance nightmares?
• Collaboration: How do you enable human engineers to work alongside agents on the same codebase?

‍

The Solution: Crafting's Infrastructure for Agents

Other teams have attempted to launch agentic infrastructures by building their own container orchestration system or by managing VM pools. But Faire's team decided to create their entire agentic stack on top of Crafting Sandboxes. This gave them several key advantages:

‍

1. Template-Based Agent Environments

Faire created separate Crafting templates for each of their distinct engineering agent workspaces (e.g. frontend, backend, data science, iOS, Android) pre-configured with the tools, dependencies, and access controls each agent type needs. When a new request comes in, a sandbox is pulled from a pool using the appropriate template. No cold start delays, no manual provisioning.

‍

2. Dynamic Sandbox Pools for Elastic Scale

Crafting's sandbox pools let Faire maintain a ready supply of warm execution environments. When demand spikes during a hackathon or major release, new sandboxes spin up automatically. When agents finish their tasks, sandboxes become available for new work or are spun down if idle. No manual intervention needed.

‍

3. Controlled Agent Access

One of Faire's biggest concerns was preventing API key leaks. With Crafting, they can provision agent sandboxes with restricted access to model API keys that human developers can't see or extract. This enables use cases like automated code review agents that run on every commit without exposing sensitive credentials.

‍

4. Multi-Cloud Support

Faire needed to migrate to a new cloud provider for AI and data-heavy workloads. Since Crafting supports multi-cloud deployments, their agent architecture remained unchanged. They simply pointed to the new cloud provider, and can fallback on their other cloud provider in case of an outage.

‍

5. Human-Agent Collaboration

Faire's engineers can choose between launching autonomous agents or collaborating directly with agents in their sandboxes. Using Crafting's shared volumes and workspace features, human developers can work in one sandbox, while an agent executes tasks in another. Both developers and agents operate on the same codebase in real-time.

‍

How Faire’s Agent System Works in Production

Faire's agent system orchestrates work requests via Slack, CLI, or GitHub comments:

1. Request Creation: A user specifies the work to be done, which agent personas to use, and which tools (via MCP configs) the agent should have access to
2. Sandbox Provisioning: A Crafting Sandbox is claimed from a pool using the appropriate template (frontend or backend)
3. Agent Execution: CLI coding agents run as a daemons in the sandboxes, with parent agents that have access to both internal tools (via a local MCP server) and external tools (via external MCP configs)
4. Task Completion: When finished, the agent executes a predefined completion script and either polls for more work or gets spun down

‍

What Makes This Approach Unique

Most companies building agentic systems end up creating custom infrastructure that's tightly coupled to their specific use cases or relying on point-solution sandboxes. Faire took a different approach by building on Crafting's sandbox infrastructure, which gave them:

• Infrastructure Reuse: The same Crafting infrastructure that Faire's engineers use for development now powers their entire agentic stack
• Ephemeral Execution: Every agent runs in a fresh environment with exactly the tools and permissions it needs
• Template-Driven Configuration: Agent capabilities and cloud resources are defined through persistent templates using Infrastructure as Code, not hardcoded logic or platform configurations
• Intelligent Orchestration: Similar requests can be routed to the same sandbox for efficiency, while maintaining isolation when needed

‍

Results: A Highly Tailored, Scalable Agentic Stack

Faire has successfully scaled their entire agentic stack on Crafting, enabling on-demand agents with access to internal systems, MCP servers, and cloud resources in a secure and scalable way. Faire is continuing to explore deeper use cases like proactive code review agents that automatically check all code before it gets pushed.

‍

Since Crafting handled execution and scaling from day one, Faire's team could pour their energy into what actually differentiates their agent system. This includes designing reusable agent personas, composable workflows, and intelligent orchestration. 

‍

Infrastructure never became a bottleneck for Faire. That allowed them to freely and rapidly iterate on agent behavior, tooling, and adoption. As a result, Faire built a mature, production-grade agentic stack much faster, instead of wasting countless hours wrangling containers and provisioning. The result is an agentic system that both works autonomously and collaborates with developers, without compromising on security or scalability.

‍

Looking Ahead

As AI agents become more capable, the infrastructure challenges companies face will only intensify. Companies that succeed will be those that can rapidly adapt to new tools and empower their teams while juggling the technical challenges of scaling agents in enterprise stacks. With Crafting, your team can focus on production-grade engineering AI workflows that multiply your velocity while we handle the underlying infrastructure for you.

‍

Faire's approach shows what's possible when you build your agentic stack on Crafting — infrastructure for teams where agents don't just assist engineers, they operate alongside them.

‍