ENGINEERING NOTES

Notes on making agent runs durable.

First-hand technical writing on crash recovery, idempotent tool execution, checkpointing, and how Actae draws the durability boundary. No product announcements — just the mechanics and the trade-offs.

13 min read

Fork a crashed agent run without re-running the prefix

A runnable walkthrough on the free self-hosted runtime: guard a side effect with the execution ledger, fork the run at the exact failing step, continue the tail, and prove it. Plus the two falsifiable measurements (fork fidelity and crash consistency) a skeptical reviewer should ask for.

forkingdurable executionagentsself-hostedeffect ledgertutorial
Read the note
12 min read

Fork the execution, not the trace

A failed agent run shouldn't be a terminal artifact. Fork any execution from an exact boundary, change one decision, and continue with the real inherited state — without replaying the expensive prefix or re-firing production side effects.

forkingcounterfactualdurable executionexperimentsagentsside effects
Read the note
9 min read

What a crash actually costs: a durability benchmark

We killed real worker processes in the middle of an external effect and counted the duplicate side effects. A durable ledger alone does not close the window — the stable operation key does.

benchmarkdurable executionidempotencycrash recoveryside effects
Read the note
11 min read

Never ran, or ran but I didn't hear back?

Replay is the easy 80% of durable agent execution. The part that bites is the side-effect boundary, and the question a log can't answer on its own: did step 7 actually finish, or did I just not hear back?

idempotencydurable executionexactly-onceside effectsagents
Read the note