AtomCode

The deciding trait is the two-command rhythm. Type /plan and it explores read-only, so you get a map before anything on disk moves; set /goal and you have told it what finished means instead of hoping it works that out. Those two habits are what separate a useful session from an expensive one, and here they are commands rather than prompt discipline you have to remember.

Against that, it is young and the crowd around it is small. Pick it when you want a light agent that asks first. Pick Aider when you want the well-worn path.

The failure mode is a confident wrong answer. Verification before declaring success is a syntax check, which catches a broken brace and nothing else, so a change that compiles and inverts a condition passes. Worse, the step budget scales with how many files were edited, so an agent that touches more of the tree buys itself more turns to keep going.

What it does right is bound the run at all. A budget exists and it is stated, which is more than most loops of this shape carry, and a stated number is a number an operator can argue with.

1. Malformed tool arguments are the most common practical failure in a tool-calling agent, and layering repair strategies is the pragmatic response to a model that emits nearly valid JSON. 2. Detecting repetition as a named mechanism means the pathology has been observed and encoded, not left to a turn limit to absorb.

2. Per-turn structured logs make the trace machine-readable, which is the difference between an anecdote and a measurement. No benchmark is published, so the design argument stands alone, but the design argument is unusually specific for a project of this size.

The interesting question is who benefits. This costs nothing and sells nothing, which means its reason to exist is showing what the hosting platform's ecosystem can produce. That is real funding of a kind, and it is also the least durable kind, because a demonstration project ends when the thing it demonstrates stops needing demonstrating.

Moat: none, and two hundred stars is not distribution. Likely path: quiet dormancy, or a maintainer picking it up personally when the sponsor moves on. Position: fine to use, wrong to standardise, and assume you inherit it if you build anything on top.

Start with distribution, because that is where this stops. There is no package name and no documented install line, which means my platform team is building and signing binaries before anyone types a prompt. Windows is not a supported platform, and a tool that covers two thirds of the estate is a tool I have to defend in two conversations.

Beyond that there is no console, no policy layer, no audit record and no contract to point a questionnaire at. Not yet: bring me a signed artefact, a Windows build and a named maintainer.

The parts I care about are right. Permissive licence, a compiled binary with no interpreter dragging its own dependency tree behind it, and one setting that decides which inference service answers. Nothing about it assumes a vendor account exists, which is the property I keep failing to find in tools with money behind them.

Then it stops. No protocol support means my servers stay on the shelf and every extension is a patch to somebody else's Rust. I can fix it myself, which is the point, but I would rather have spent that evening on something else.