Chidori

The trait that decides it is patience. A run can suspend to disk while it waits for an answer and resume days later in a fresh process, so a workflow with a human in the middle stops being a queue, a database and three cron jobs you wrote yourself. That is a genuinely large amount of plumbing you do not build.

What you accept is a small project and a young one. Pick it if waiting is part of your problem shape. Pick Julep if you would rather somebody else operated the durability.

Two costs come with the design. Running a pure-Rust engine instead of the standard runtime means native modules and much of the ecosystem simply do not load, which a team discovers when a dependency fails rather than when they choose the tool. The second is that a complete record of every host call includes credentials and payloads, and a log that valuable needs handling nobody documents.

What it does right is refusing by default. The policy denies capabilities unless injected, with resource limits, which is the correct direction for generated code.

1. This is the strongest reproducibility property on the board. Because every side effect passes through the runtime as a recorded host call, a past execution is fully determined by its log, and replay requires no provider, no network and no spend.

2. The consequence is that regression testing an agent becomes ordinary software testing, which is the thing this field has been unable to do. 3. No published evaluation exists, and none is needed for the claim, because a reader can verify it by running a checkpoint twice.

The technology is more defensible than the company, which is the wrong way round. Durable execution is a category where the incumbents sell managed services, and this gives away the hard part with nothing beside it, so there is no revenue to fund the engineering that made it interesting.

Moat: the runtime design, which is real and unmonetised. Likely path: an acqui-hire by a workflow or agent-platform vendor that wants the replay mechanism, which is the outcome I would bet on. Position: adopt it, keep your checkpoints, and assume the authors end up somewhere else.

Testing agent behaviour has been the thing I could not budget, because every run costs tokens and nondeterminism makes the result meaningless. A recorded execution replayed in our automation changes that arithmetic completely: the check costs compute minutes we already buy, and it fails when behaviour changes.

What is missing is a supplier, a support commitment and any operational tooling, so my platform team owns it entirely. Onboarding is a week. Approved with conditions: one service, our own log storage for the recordings, and a named owner.

No language runtime to install is the detail that sells it. One compiled artefact, and the scripting layer lives inside it, so deploying this to a machine is copying a file rather than managing a version manager and a lockfile. Three install paths exist and all of them end at the same binary.

Pointing inference at my own server is one environment variable, named in the documentation, so the whole thing runs on hardware I own. Permissive licence over Rust I can read. Very little here annoys me, which is unusual.