Jukebox-OS brings deterministic time-travel replay and ACID sandboxes to AI agents

An agent runtime built like a database

An open-source project called Jukebox-OS reached the Hacker News front page on October 25, according to the site's feed. The project is a runtime for AI agents that borrows two ideas from database engineering — a write-ahead flight recorder and ACID transactions — and applies them to the problem of agents that mutate the outside world while they work.

According to the project's README on GitHub, the core is written in Rust with Python bindings and sits underneath common agent frameworks including LangGraph, CrewAI, AutoGen and the OpenAI and Anthropic SDKs. Marius Carlsson is credited as the architect. The stated motivation is a failure mode many agent operators will recognise: a multi-step workflow that edits SQL tables and local files partway through, then derails several steps later, at which point checkpointing the conversation is useless because the real-world changes have already happened.

Everything recorded, everything hashed

The first half of the system is a recorder. Every LLM request, streamed token chunk, random seed, tool input and output, clock tick and filesystem mutation is hashed with BLAKE3 and BLAKE2b and appended to a portable binary tape file with a .jbx extension. Because the tape is content-addressed, the project claims a 50-step agent run can be replayed offline in roughly 0.4 milliseconds with no API cost, and that a tape from a failed run can be attached to a bug report the way engineers attach core dumps. An OpenTelemetry exporter for GenAI spans and metrics is included as well.

Side effects held in escrow

The second half is the sandbox, which the README names the Shadow Booth. File writes, shell commands, SQL queries and MCP tool calls are intercepted in a copy-on-write WASI staging layer and held there until surrounding code explicitly commits. If a verifier fails, a rollback is described as leaving nothing behind on the real system. The quickstart example shows an accounting agent staging a refund to a ledger file and rolling itself back when the amount passes a $500 autonomous threshold — the kind of policy gate that usually has to be bolted on afterwards.

Forking timelines and tripping breakers

Tapes can also be forked. A developer loads a failed trace, jumps to a chosen frame, swaps the model or rewrites the system prompt, and resumes from that point. Steps before the fork resolve from the tape; only steps after it touch the live network, and the two timelines can be diffed side by side. A CLI lists stored recordings, though the README text available in the feed cuts off partway through that section.

A circuit breaker rounds out the runtime. It watches streaming traffic between agents using semantic similarity and AST diffs, tracks token spend velocity, and trips automatically when agents fall into a repetitive loop or when a per-track cost cap is hit — the README's example sets that cap at $0.50.

Claims are the project's own

Every figure above comes from the project's documentation, which is the only material available here; the performance numbers have not been independently benchmarked, and the feed says nothing about adoption or outside review. The repository lives under a personal GitHub account and the README leans heavily on diner theming, so teams will want to read the code before pointing it at production systems.

Why it matters

Determinism is the missing ingredient in agent debugging. Model outputs are probabilistic and tool side effects are persistent, and that combination makes agent failures nearly impossible to reproduce — incidents get investigated by rereading logs and guessing. Record-and-replay tamed similar problems in databases, in distributed systems through event sourcing, and in low-level debuggers. Applying the same discipline to agents, where any failure becomes a shareable artifact and any rollback actually undoes the damage, is an obvious next step, and transactional side effects lower the barrier to letting agents near production data at all. Whether Jukebox-OS becomes the implementation teams standardise on is an open question, but the pattern it encodes — record everything, stage effects, commit on verification, fork to investigate — looks like the shape credible agent infrastructure will take.