Apache Maka

The deciding trait is that the interfaces are interchangeable. Desktop, terminal and command line are thin clients over one runtime, so the session you started in a window is the session you resume in a shell, and nothing has to be exported or reconstructed to move between them.

That sounds like a small thing until you have lost an hour of context by switching windows. What you should know before starting is that this is early software and it feels like it. Pick it if you move between surfaces. Pick a plain terminal agent if you never have.

The failure mode is growth. Every model message, every tool call and every permission decision is written and never deleted, which is exactly the property that makes recovery work and exactly the property that makes a directory unbounded. Long sessions with large tool output are the expensive case.

Nobody notices this on a laptop for a month and everybody notices it on a shared machine in a quarter. No compaction of the record, no rotation and no documented cleanup exist. What it does right is the honesty of the design: a record that can be trusted is worth more than one that is small.

1. Holding the model constant and varying the harness is the right experiment, because it isolates the variable the product actually controls. Most published numbers confound the two and are therefore uninformative about the tool. 2. Using the official verifier rather than a self-authored one closes the second common escape.

2. Per-task publication is the third good decision: an aggregate hides which tasks a harness fails, and a reader comparing harnesses cares about precisely that distribution. The result is a set of numbers a sceptic can argue with, which is a lower bar than it should be and one almost nobody clears.

Incubation is a governance answer rather than a funding one. The project cannot be acquired and cannot be redirected by an investor, which removes the two ways these usually end, and replaces them with a slower question: whether the contributors who brought it in stay interested through graduation.

That question at least has a public answer, which is more than most rows here offer. Moat: neutrality, plus a name that carries weight with the people who buy nothing. Likely path is graduation or quiet dormancy. Position: structurally the safest bet in this category and the least finished one.

I cannot deploy an unsigned binary to sixty managed laptops. That is not a preference, it is an endpoint policy written long before any of this existed, and the exception process costs more attention than the tool would save in its first quarter of use.

Everything else reads well. The licence is free, the sessions stay on the machine, and it runs unattended, which means it could become a measured step later. Not yet: bring me a signed, versioned release and I will have this conversation properly instead of stopping at the download page.

The boundary is the part I did not expect to like. Tools that write files or run a shell cross an explicit line first, which means the dangerous half of an agent has a named edge rather than a permission dialog I stopped reading in week two of using it.

The licence is permissive, the model is whatever I point it at including one on my own hardware, and nothing phones anywhere by default. What I do not get is the protocol, so the servers I already run stay outside. Grudging respect for a project that got the security model right first.