The host is written in Rust, with `extern "C"`, which makes it able to be loaded as a C library by programs written in other languages. Most languages have support for this.
It's also designed to be run in an event loop. I've tested this with Bun's event loop that runs TypeScript. I haven't tried it with other async runtimes, but it should be doable.
As for the browser, I haven't tried it, but you might be able to compile it to WASM -- the async stuff would be the hardest part of that, I suspect. Could be cool!
I generally agree. TypeScript is a great language, and JS runtimes have certainly had a lot of money and effort poured into them for a long time. I would add WASM to this category, as probably the closest thing to Mog. Write a program in some language, compile it to WASM, and load it into the host process. This is (probably) nice and safe, and relatively performant.
Since it's new, Mog will likely not yet beat existing systems at basically anything. Its potential lies in having better performance and a much smaller total system footprint and complexity than the alternatives. WASM is generally interpreted -- you can compile it, but it wasn't really designed for that as far as I know.
More generally, I think new execution environments are good opportunities for new languages that directly address the needs of that environment. The example that comes to mind is JavaScript, which turned webpages into dynamically loaded applications. AI agents have such heavy usage and specific problems that a language designed to be both written and executed by them is worth a shot in my opinion.
JIT means the code is interpreted until some condition kicks in to trigger compilation. This is obviously common and provides a number of advantages, but it has downsides too:
1) Code might run slowly at first.
2) It can be difficult to predict performance -- when will the JIT kick in? How well will it compile the code?
With Mog, you do have to pay the up-front cost of compiling the program. However, what I said about "no process startup cost" is true: there is no other OS process. The compiler runs in process, and then the compiled machine code is loaded into the process. Trying to do this safely is an unusual goal as far as I can tell. One of the consequences of this security posture is that the compiler and host become part of the trusted computing base. JITs are not the simplest things in the world, and not the easiest things to keep secure either. The Mog compiler is written entirely in safe Rust for this reason.
This up-front compilation cost is paid once, then the compiled code can be reused. If you have a pre-tool-use hook, or some extension to the agent itself, that code runs thousands of times, or more. Ahead-of-time compilation is well-suited for this task.
If this is used for writing a script that agent runs once, then JIT compilation might turn out to be faster. But those scripts are often short, and our compiler is quite fast for them as it is in the benchmarking that I've done -- there are benchmarking scripts in the repo, and it would be interesting to extend them to map out this landscape more.
Also, in my experience, in this scenario, the vast majority of the total latency of waiting for the agent to do what you asked it is due to waiting for an LLM to finish responding, not compiling or executing the script it generated. So I've prioritized the end-to-end performance of Mog code that runs many times.
Did somebody say 'global namespace'? I spent years working on one of those as part of Urbit... In general, I think you're right. Each conversation is an append-only log at the lowest layer, and I see no reason not to expose that fact as a global namespace, as long as permissions are handled gracefully.
Of course getting permissions to work well might be easier said than done, but I like this direction.
Hi, I'm the other author on this paper. You've asked a good question. I had originally planned on writing an agentic_reduce operator to complement the agentic_map operator, but the more I thought about it, the more I realized I couldn't come up with a use case for it that wasn't contrived. Instead, having the main agent write scripts that perform aggregations on the result of an agentic_map or llm_map call made a lot more sense.
It's quite possible that's wrong. If so, I would write llm_reduce like this: it would spawn a sub-task for every pair of elements in the list, which would call an LLM with a prompt telling it how to combine the two elements into one. The output type of the reduce operation would need to be the same as the input type, just like in normal map/reduce. This allows for a tree of operations to be performed, where the reduction is run log(n) times, resulting in a single value.
That value should probably be loaded into the LCM database by default, rather than putting it directly into the model's context, to protect the invariant that the model should be able to string together arbitrarily long sequences of maps and reduces without filling up its own context.
I don't think this would be hard to write. It would reuse the same database and parallelism machinery that llm_map and agentic_map use.
WASM is a great system, but quite complex -- the spec for Mog is roughly 100x smaller.