An AI agent can generate a code change in seconds. That change touches a database schema, a caching layer, an auth service. Testing it in a sandbox tells you whether the code compiles, which is a long way from telling you whether it works.
Sumeet Vaidya built Crafting to close that distance.
Vaidya was one of two initial engineers rebuilding Facebook's Groups product and contributed to early React Native releases. He went on to manage engineering at Uber across international growth, vehicle solutions, and micromobility. During that stretch, an internal team built a chat tool called U Chatt to replace HipChat. The engineering was serious. The reliability problems were unresolvable. Uber moved to Slack. The lesson Vaidya carried forward was about trust. "Trust is everything when it comes to enterprise products," he's said. Once reliability breaks, recovering it is a fundamentally different problem than building it.
Later, directing engineering for bots at Discord, he watched third-party code run against platform infrastructure at scale. Each of these roles involved the same underlying reality: as services multiply and interact, the distance between "this code looks right" and "this code works in production" widens. For human engineers, that distance was already painful. For AI agents generating changes at speed, it becomes something closer to a wall.
"We are speeding headfirst into a massive brick wall of AI-generated code, and the old methods of sandboxing and manual review simply aren't going to cut it anymore."
Crafting for Agents, backed by a $5.5M seed round led by Mischief with WndrCo, gives agents controlled access to real production dependencies without exposing production itself. The platform maps a company's service topology and spins up only the relevant services for each change, with scoped credentials and environment-level access controls. Agents write code, test against real infrastructure, see what broke, fix it, iterate.
The product provides a controlled window into the actual dependency graph, with enough isolation to be safe and enough fidelity to be meaningful. Think of it as a live cross-section of production, scoped to exactly the services a given change will touch.
Faire, the wholesale marketplace, shows what the problem looks like before that window exists. When their team tried building an agentic system to autonomously write and modify code across their codebase, three problems surfaced immediately: giving agents access to internal systems without risking credentials, provisioning environments for unpredictable workloads, and supporting diverse toolchains without creating maintenance overhead. Before Crafting, they were stitching together point solutions and spinning up one-off environments manually.
The infrastructure Crafting built for AI agents is the same infrastructure Faire's human engineers use for development. Agents inherited the validation problem and accelerated it. The cost of skipping validation grew with the speed.
That continuity is also what makes Vaidya's bet interesting and genuinely unproven. Crafting started as a cloud development environment for human engineers. The pivot toward agents as the primary user is recent. Whether the validation layer that human engineers needed translates cleanly to agents operating at a different speed and scale remains the open question embedded in the product. Vaidya has described Crafting's origin as addressing "challenges we were experiencing ourselves as engineering leaders." The product reflects that directness. It gives agents access to the thing that tells them whether they're right.
Things to follow up on...
- The cancellation forecast: Gartner predicts more than 40% of agentic AI projects will be cancelled by end of 2027, not because the technology failed but because the foundation underneath it was never right.
- Quality as production killer: A LangChain survey of 1,300+ practitioners found that 89% have implemented observability for their agents but only 52% run systematic evaluations, suggesting most teams can see what's happening but can't validate whether it's correct.
- Compounding accuracy math: If an AI agent achieves 85% accuracy per action, a 10-step workflow only succeeds about 20% of the time, a dynamic explored in detail in Composio's report on why agent pilots fail in production.
- Shadow agents emerging: As MCP adoption scales past 97M monthly SDK downloads, security teams are watching for "Shadow Agents" running unvetted on developer laptops with access to critical data systems.

