The event log: replay-exact, restart-safe
Agents that work for hours need a memory that survives anything. Muse Code’s runtime is built around a single append-only log — and almost every trust property of the product falls out of it.
One log, everything in it
The design is almost aggressively simple: Muse Code appends every model call, every tool run, every approval, and every edit to a local event log. Not a chat transcript, not a summary — the actual sequence of operations, in order, on your machine. Meta calls it the runtime’s single source of truth, and the phrase is literal: whatever question you have about what an agent did, the log is where the answer lives.
Restart-safe: crashes stop being expensive
Because the log records everything up to the moment of failure, a crash doesn’t reset the session — the agent resumes precisely where it stopped, without re-running prior steps. For five-minute tasks this is a nicety. For the long-horizon work Muse Code is aimed at, it’s the difference between the product working and not: Meta’s kernel-optimization case study ran 1,000+ tool calls over as much as 24 hours. Nobody runs a 24-hour job on a runtime where hour 23 can evaporate.
Replay-exact: audit by stepping through
The same property that makes restarts safe makes sessions reviewable. muse replay walks a recorded session step by step — every subagent spawn, every steer, every cancel, in the order they happened. When an agent made a decision you don’t understand, you don’t interrogate it; you replay it and watch.
Full traces also export. That turns an agent session into something you can hand to a teammate for review, or attach to a compliance record — an outsider gets enough context to evaluate the agent’s reasoning without re-running anything. Teams that treat AI-generated changes with (justified) suspicion get a code-review artifact for the process, not just the diff.
Why this matters more with multiple agents
A single agent’s history is easy to eyeball in a terminal scrollback. Muse Code runs persistent background agents and fans work out to parallel children — activity you are, by design, not watching in real time. “Fully auditable” is the counterweight that makes “multi-agent by default” acceptable: everything those agents did while you weren’t looking is transparent, traceable, and replayable after the fact. Autonomy and auditability arrive as a pair, not a trade-off.
The bigger pattern
Event sourcing — deriving state from an append-only log of facts — is a decades-old idea from distributed systems, and it’s quietly becoming the standard answer to agent reliability. What Muse Code demonstrates is how much product surface one primitive can carry: crash recovery, step-through debugging, session handoff, and compliance export are all the same log wearing different hats. Start with a small task and replay it — the quickstart takes about five minutes.