Agent Executor (AX)

The trait that decides it is recovery. An interrupted run picks up where it stopped instead of starting over, which for hour-long tasks is the difference between a platform and a demonstration, and it works across a distributed deployment rather than only on one host.

This is infrastructure, not a tool: you will not open it in the morning, your platform will. Pick it if you are the team providing agents to other teams. Pick Container Use when one isolated environment per agent is all you actually needed.

Two facts compound. Interfaces are declared unstable, and the usual remedy, contributing the fix or the compatibility shim yourself, is closed because outside contributions are not being accepted. That combination means every upgrade is a migration performed on somebody else's schedule with no channel to negotiate it.

What it does right is scoping the blast radius. Each actor brings its own environment, tools and instructions, so an experiment that goes wrong is confined to one actor rather than to the controller.

1. Two design choices carry the system. Suspendable images make a run's state a first-class artefact, so resumption is restoration rather than replay, and the token cost of a recovered run is close to zero. 2. A single writer for state avoids the consensus problem entirely instead of solving it badly, which is the correct trade at this scale.

2. No evaluation is published: no recovery success rate, no overhead figure for suspension, and no comparison against restart-from-scratch.

Funding is not the risk here and never will be. The risk is that infrastructure released by an organisation this size exists to serve an internal strategy, and internal strategies are revised without a press release. Projects in this position are either absorbed into a commercial platform or archived, and both outcomes look the same until the day they do not.

Moat: the sponsor's own agent ambitions, which this supports rather than monetises. Likely path: a managed service built on it. Position: adopt the pattern confidently, adopt the repository cautiously.

Unattended execution is what makes this relevant to me at all: work can be scheduled and recovered without a person watching, which is the property my platform team keeps asking for. Licensing is free and the binary is a single install.

The supplier side is the problem. There is no support commitment, no stability guarantee and no route for my engineers to submit a fix upstream, so any patch we need becomes a fork we maintain. Onboarding a platform engineer is a week. Approved with conditions: a pinned version, one internal workload, and an owner for the fork we will end up keeping.

Per-actor tool configuration is the detail I like: my servers attach to the specific actor that needs them rather than to a global registry everything shares, which means capability is scoped the way it should be. A single compiled binary from the language's own package tool is the right install story for infrastructure.

The permissive licence keeps a fork legal, which matters more than usual given contributions are closed. What I do not get is the model layer, where nothing local is recorded, so inference stays remote.