The Setup

EthGlobal Cannes 2026 was a special one. Set against the Mediterranean, with 600+ hackers from across the world, the atmosphere was electric. For us, it was our second EthGlobal event after New York 2025, so we had some idea of what to expect. But every hackathon is different.

We went in with a thesis: AI agents are increasingly controlling wallets and executing transactions autonomously, but there's no security layer between the agent's decision and the blockchain. We wanted to build that layer. ENShell was born from that idea.

Building Under Pressure

The first 12 hours were about architecture. We mapped out the four-layer design: the SDK for encryption, the smart contract for policy management, the Chainlink CRE workflow for confidential analysis with Claude, and the human-in-the-loop escalation via Ledger.

By hour 18, we had a working prototype that could intercept an agent's transaction, encrypt it, send it through the CRE for analysis, and return a verdict on-chain. The biggest technical hurdle was getting encryption to work inside the CRE runtime, which doesn't support traditional JavaScript crypto libraries. We had to find creative workarounds for that.

Scope management was tough. Every hour, someone on the team wanted to add one more feature. Knowing when to stop adding and start polishing is something you only learn by doing.

What Worked

Starting with the demo. We wrote the demo script before we wrote the code. This forced us to think about what would actually resonate with judges, not just what was technically interesting to us.

Chainlink CRE as the backbone. Using Chainlink's Compute Runtime Environment meant we could run Claude on encrypted agent instructions without ever exposing plaintext on-chain. This was the key architectural decision that made ENShell possible in 36 hours.

ENS for identity. Instead of building our own agent identity system, we leveraged ENS subdomains for agent registration and reputation scoring. This gave us composability with the broader Ethereum ecosystem without reinventing the wheel.

What We'd Do Differently

We spent significant time building a Ledger Live companion app for hardware wallet approvals. While the concept is strong (physical approval for escalated agent actions), the reality is we didn't have enough time to properly showcase it to judges. Sometimes less is more. That time would have been better spent deepening the core CRE workflow and preparing a tighter demo. The Ledger integration added complexity to the QA process without delivering proportional impact during the presentation.

The Result

Finalist status and the Best Workflow with Chainlink CRE prize. We validated that an on-chain firewall for AI agents is a real, buildable product that solves a problem people recognize immediately.

ENShell is now live at enshell.xyz and we're continuing development beyond the hackathon.

Takeaways for Builders

  1. Ship the demo first. Work backward from what you want to show on stage.
  2. Use existing infrastructure. Chainlink, ENS, and Ledger each saved us from building things from scratch.
  3. Cut features early. A focused MVP with a clean demo beats a sprawling prototype every time.
  4. Get the architecture right first. The first few hours define everything. Invest them in design, not code.