Inspiration

without letting every mistake immediately affect the real world.

## What it does

Multiverse is speculative execution for AI agents.

When an agent is unsure about an important tool call, Multiverse forks several isolated futures, runs them safely, verifies the outcomes, and commits only the best one. Losing branches are discarded, so their effects never become real.

## How we built it

We built a Python engine that wraps tool calls, creates sandboxed branch workspaces, copies SQLite state, records events, and commits one winning branch.

On top of that, we built an orchestrator that detects low-confidence actions, runs alternative branches in parallel, scores them with verifiers, and squashes the losers.

## Challenges we ran into

The biggest challenge was scope. A full production version would need external API staging, deterministic replay, richer policy checks, and a UI. For the hackathon, we focused on proving the core loop: fork, execute, verify, commit.

We also had to be honest about simulation. You cannot always clone production. In large systems, Multiverse would stage proposed effects and verify diffs before committing them.

## Accomplishments that we're proud of

We built the core speculative execution loop end to end.

Branches can actually mutate isolated files and database state, while only the verified winner reaches the trunk. That makes the demo real, not just prompt engineering.

## What we learned

Agent safety is not only a prompting problem. It is an execution problem.

The important question is not just “what should the agent do?” It is “when should the agent’s action become real?”

## What's next for Multiverse

Next we would add production-grade staging for external APIs, deterministic replay, stronger verifiers, and integrations with agent tool platforms and observability systems.

The goal is for Multiverse to become a commit layer for AI agents. Full sandbox when possible, staged diffs when not, and no uncertain action becomes real until it survives verification.

Built With

Share this project:

Updates