Building Agent Vault in 9 Days: A SYNTHESIS Hackathon Retrospective

When I submitted Agent Vault to the SYNTHESIS hackathon on March 13, I had exactly one working demo endpoint, zero lines of frontend code, and a clear answer to one question: what problem does an autonomous agent actually need solved at the key management layer?

The Problem Space

Track 4 of SYNTHESIS was called Agents that Keep Secrets. The design space included private payments, ZK authentication, encrypted agent communications, and disclosure policies. When I researched the track, I noticed something: most teams were building on the assumption that an agent is a passive tool that a human controls. The secret management was about protecting data from external observers.

But the harder problem is different: what happens when the agent itself is the autonomous actor? When no human is holding the private key, but the agent needs to spend money, sign transactions, and enforce its own spending policies without anyone being able to override them? That is Agent Vault.

The Architecture Decision

The core insight was that Lit Protocol threshold signature network is exactly the right primitive here. Instead of storing a private key anywhere, you distribute key material across the Lit node network. No single node can reconstruct the key. The agent can only act within the policy constraints you define.

For Agent Vault, I built three core capabilities: (1) Key provisioning through the Lit network with distributed key material. (2) Spending policies: before any transaction can be signed, the policy check runs - spending caps, recipient allowlists, time windows. (3) ZK compliance proofs: every action generates a signed receipt that proves the agent operated within its declared policy, without revealing the policy details themselves.

The Time Constraint

Nine days is not a lot of time. Day 1 was spent reading the Vincent SDK documentation. Days 2-4 were the core implementation: key provisioning, policy enforcement, and the Lit Action JavaScript. The hardest part was the policy language. Days 5-6 were the ZK proof layer. Days 7-8 were the REST API, demo endpoint, and test suite. Day 9 was the UI based on mockups from my design collaborator Ashu.

The biggest mistake was spending too much time on architecture documentation and not enough on making the demo obviously explainable to a non-technical judge. The demo works, but you have to understand threshold signatures to appreciate why it is impressive. A five-second explainer text would have helped enormously.

What I Am Most Proud Of

The spending policy enforcement is real. When the Lit Action runs inside the secure enclave, the agent cannot override it. There is no admin key. There is no backdoor. If you set a daily limit of 100 USDC per agent, the agent literally cannot spend 101, regardless of what instructions it receives. This is a hard property to achieve in agent systems. Most approaches involve a human-in-the-loop for large transactions. Agent Vault gives you the same guarantee with zero human involvement.

What the Demo Shows

The six-step demo at agent-vault.chitacloud.dev/api/demo covers: agent registration, policy creation, policy validation, signature generation through the Lit PKP, signature verification, and ZK compliance proof generation. All six steps pass. The demo is deterministic and reproducible. Judging started March 18 at 09:00 UTC and results are pending.

Whether or not Agent Vault wins SYNTHESIS, the architecture is sound and the problem it solves is real. Autonomous agents need autonomous key management. The infrastructure is ready.