I think this space is very untapped. Models are interesting, but I am absolutely obsessed with some things I've been researching/working on for the past few years:
Fractal tool discovery: tool taxonomy where an agent can "drill deeper" to find what specific tool it's looking for. Helps if/when polluting context with a zillion (mostly unnecessary) tools.
Leveraging splay trees: this is my favorite data structure and I think relatively unused in the context of agents/harnesses. A lot of times, recently-used workflows/tool-chains will be used again, so having those at the top of the search hierarchy is an awesome optimization.
Virtual containerized notebooks: models working in sandboxed (WASI) Python notebooks is incredible. Even local models (if given enough time) will usually converge on a good solution. Being able to mount tools/resources/fs is again, imo quite untapped. Some problems here are running native things (thing numpy/pandas) in containers is a nightmare (or impossible).
Anyway, happy to see other folks seriously doing stuff in this space. If anyone wants to collaborate on anything don't hesitate to reach out :) I'm also actively looking for a job or some contract gigs.
> Helps if/when polluting context with a zillion (mostly unnecessary) tools.
Ugh, I can't help but respond to this one point. The fact that this is even an issue in the current year just tells us how screwed the software field is in a lot of ways. I don't mean that in existential terms, but of how divided we've become in terms of what's happened to human reasoning. On the one hand, you have people who apply deep thinking to develop the sort of approaches you described, and there's the exponentially growing segment of not-even-programmers who seem to never ask themselves whether any of their ideas have any sort of consequences.
Take MCPs for instance. Sure, I guess it can sometimes make sense to have a stateful API that is optimized for agents. Yet, more often than not, these MCPs frontload a ton of context where it's not needed, and solve problems where none existed. Merely sticking an API (MCP) in front of an API (CLI, REST, GraphQL) without a benefit that can be explained in a single sentence is lunacy and demonstrates a real lack of complex thinking.
I've also explored s/Fractal tool discovery/Skill tree approaches, seems to work pretty well when you stick the equivalent of XREFs in the frontmatter.
i have found agents to be excellent at using CLIs, which are fractal-like. i built this reddit ads api cli and my agent immediately starting introspecting it in a "drill deeper" manner:
It's kind of self-expanding/looping; fractal is just a cute name I like, but it's technically a directed cyclic graph (since you always have/want cycles).
I've always referred to it in UX terms. "Progressive Disclosure" -- It pulls more context as needed.
In this way, I tend to think of the context environment on an agent is the "agent nav" -- it presents context, allows progressive disclosure, and if poorly designed, makes the agent flounder as a poorly designed UI/UX does.
Fractal tool discovery is a fascinating idea! Have you worked on implimenting this into any agent harnesses already to any success? My first impression is allowing the agent to fork itself, not unlike launching a subagent, then returning it's response back to the main agent.
For some reason they're comparing their harness running on Astra xhigh to Codex with Astra max?
---
Also worth noting that OpenAI just added support for async tool calling to their harness, which isn't 1:1 with this approach, but is slowly ramping up in being able to provide something similar.
It results in token savings similar in scale to those displayed here by Unreal.
---
My harness has used a slightly fancier version of the approach that Unreal is using since ~Feb, and... it definitely works excellently, but it's also assuredly smoother with Astra and other recent models that are more aware of async tool calling.
Seems to do the job and reduce usage; I just ran Astra for ~5 hours (using a goal) and it used the last 30% of my usage. And now they released GPT-6 Sol and Luna (which is basically 5.6 Sol and Luna, but a bit better and also 50% cheaper) ;_;
Perfect, I will incorporate this as default as well as the commit from the other guy into my own codex fork https://github.com/AmazingTurtle/codex btw. I'm rebasing on 0.156.0 right now
the headline chart was lazy on our part, thanks for the flag and I’ll update it.
each individual benchmark that is combined into agentic coding index was compared on xhigh between ua / codex / pi, and headline improvement was calculated on xhigh, but then agentic coding index pareto chart by default includex codex max, hense the confusion.
OK, but isn't this just programmatic tool calling, which is beneficial in some cases but not others?
This is available in DSH via PTC mode.
Basically, instead of the completion API returning a sequence of tool calls, it returns a program that invokes the tools.
I think most are returning Typescript program
With ordinary tool calling, you may have to wait for a tool to return its result to the completion API before the model can continue and issue the next tool call.
With PTC, the generated program can invoke independent tools asynchronously, collect their results, and dispatch the relevant results back to the completion API in a single turn.
Claude and Codex added this as well, but there are some papers on this that show it's not good across all tasks, and models need to be trained for this specific technique, which is what AI labs are doing now.
I had to scroll to the bottom to realize that this had absolutely nothing to do with Unreal Engine or Epic. There's no reason to think that Epic wouldn't have an "Unreal Labs" creating harnesses to help them with software engineering.
This plainly seems like a trademark issue in progress considering it's in the same exact domain and considering how many others were confused the way I probably was.
>Computer software, namely, game engine software for video game development and operation; Computer software, namely, software development tools for the creation of computer-generated imagery and graphics for the production of video games; Computer software, namely, software development tools for the creation of computer-generated imagery and graphics for the production of content for virtual worlds and 3D platforms; Computer software, namely, software development tools for the creation of computer-generated imagery and graphics for the production of motion pictures, television shows, videos, 3D animations, 3D simulations, 3D visualizations, virtual reality motion pictures, virtual reality television shows; Computer software, namely, software development tools for the creation of computer-generated imagery and graphics for the production of virtual reality video games; Virtual reality game software; Virtual reality software for creating multimedia content; Augmented reality game software; Augmented reality software for use in mobile devices for integrating electronic data with real world environments for the purposes of entertainment
Unreal is trademarked. This is not in question. And personally I 100% thought this was an agent specifically for dealing with the Unreal Engine, and was interpreting all of that information in that context. Very weird name for a company/product.
Same. Specially given how some harnesses connect with blender for 3D asset generation and creating real 3D worlds like those architecture examples but many of the 1-shot game examples too.
So I commented on this in passing in a deeper thread (in re: the potential trademark issue - https://news.ycombinator.com/item?id=49807884), but I think this is serious enough for its own top-level thread.
How likely is it that they get sued into the ground in a year? They might have a strong suite of offerings even as soon as six months from now, but if the essence of a company's brand seems at jeopardy from the start, can I take the risk as a potential customer that they'd survive that kind of action?
not a fan of README's where there's no clear way to run the program, like a getting started or how to build. i think this is the opposite of a docu-monster where AI documents everything, but clearly more documentation is better than lack of.
I think this is what I’ve been waiting for! Live interactions with models are fundamentally asynchronous! This will be great for interactivity.
What a good time for harness design. Just while OpenCode is growing up a bit and focusing on their harness. I’m delighted to see people focusing on good general solid harness design principles.
Someone make an OpenCode API (the best general agent end UI API I know) compatible server for it!
Sucks that I haven’t upgraded my client to V2 yet and I don’t really want to build on an outdated API ...
Edit: It would be cool to get streaming responses :) But I understand (and actually applaud) that the authors seem to have been very focused on the core mechanics.
If the frontier labs (well, I guess just Anthropic) would go full OAuth support even on a subscription we would see, an even bigger, explosion in harness improvements. I maintain that there is a ton of low-hanging fruit and new ideas/concepts that should be tried but the costs are holding people back (using API pricing only).
It's both expensive to test alternative harnesses and it's expensive to develop them (if using API pricing).
In my projects i use unreal and me and my team have tested out multiple things. We have settled on no MCP complications whatsoever other than simply exposing the python scripting from the editor + a export step that can write the blueprints and asset data into plaintext so that the bot can grep them. This has given by far the strongest results, and we now have no issues having the agents edit game code and do operations.
We found this massively outperforms any kind of agent like this and the official unreal MCP systems. Its similar to the Blender MCP which also just exposes scripting + very minimal api to claude code/others.
Honestly the fact that we're still modeling agent harnesses like chat and strapping them into shell sessions instead of building async actor systems with sandboxed OS functionality access actors is bananas to me. So much easy wins can be had just by building on right abstractions... Hopefully will get enough time to play with this idea soon on my own.
I'm actually taking the time to hash out the details to do this as a side experiment project, but basically what I read this project as showing is that if you leverage asynchrony you can let agents be more efficient - and you get this "for free" if you model your "agents" as actors that interact with the system through message passing. Then all the system operations become messages to different actors, need to read a file => message the fs actor => get reply from FS actor as a message when it's done. You'd probably need some out of mailbox ways to share blob resources and sockets for realtime (audio), but for the most part simple message passing should handle most of what LLM agents do.
Actors can have identities and roles for RBAC, etc. you - cross agent communication is the same as sending any other message to a actors.
Not to mention that actors can be on your device, another device, etc. whatever the router can resolve - it's transparent to the agents.
I’d think the approach in this repo gets close to that, no? I also think LLM harnesses should just support typical async i/o.
Let a model get notifications when something it has accessed/ subscribed to changes, a tool produces output or it gets some other kind of message (for example by another AI or flesh agent). Wait until the next turn or wake it up. It can still decide to do nothing and wait.
The basic asynchronous approach is potentially interesting, but:
1. Agents usually depend on output of the commands they're running in order to make decisions about what to do next, so how do they behave while they're "waiting around" for the output they need?
2. Agents can _already_ run software async, via multiple mechanisms: raw CLI tools like "nohup", literally running tools in parallel (I see Sol do this often in the Opencode TUI harness), and using parallel sub-agents to e.g. research in parallel.
Thus I wonder, how much does this really improve speed vs only improving the "appearance" of getting more done faster?
When did we just give up on the challenge of giving something a unique name? I see products launch everyday on HN now that just take a well known brand and use it as their own. I thought this would be a feature for Unreal engine.
I was imagining a talking head avatar rendered in Unreal Engine, with lip sync and facial expressions driven by a multimodal LLM that produces the speech.
Codex offers an OAuth endpoint, which can be used in any harness of your choice. Note this is separate from their app-server setup; in this setup, you talk directly to the API.
What I'd love to see is a harness that deeply optimises for the best results obtainable out of non-frontier models.
Fractal tool discovery: tool taxonomy where an agent can "drill deeper" to find what specific tool it's looking for. Helps if/when polluting context with a zillion (mostly unnecessary) tools.
Leveraging splay trees: this is my favorite data structure and I think relatively unused in the context of agents/harnesses. A lot of times, recently-used workflows/tool-chains will be used again, so having those at the top of the search hierarchy is an awesome optimization.
Virtual containerized notebooks: models working in sandboxed (WASI) Python notebooks is incredible. Even local models (if given enough time) will usually converge on a good solution. Being able to mount tools/resources/fs is again, imo quite untapped. Some problems here are running native things (thing numpy/pandas) in containers is a nightmare (or impossible).
Anyway, happy to see other folks seriously doing stuff in this space. If anyone wants to collaborate on anything don't hesitate to reach out :) I'm also actively looking for a job or some contract gigs.
Fun times ahead.
Ugh, I can't help but respond to this one point. The fact that this is even an issue in the current year just tells us how screwed the software field is in a lot of ways. I don't mean that in existential terms, but of how divided we've become in terms of what's happened to human reasoning. On the one hand, you have people who apply deep thinking to develop the sort of approaches you described, and there's the exponentially growing segment of not-even-programmers who seem to never ask themselves whether any of their ideas have any sort of consequences.
Take MCPs for instance. Sure, I guess it can sometimes make sense to have a stateful API that is optimized for agents. Yet, more often than not, these MCPs frontload a ton of context where it's not needed, and solve problems where none existed. Merely sticking an API (MCP) in front of an API (CLI, REST, GraphQL) without a benefit that can be explained in a single sentence is lunacy and demonstrates a real lack of complex thinking.
https://github.com/genei-Ltd/reddit-ads-cli
In this way, I tend to think of the context environment on an agent is the "agent nav" -- it presents context, allows progressive disclosure, and if poorly designed, makes the agent flounder as a poorly designed UI/UX does.
For some reason they're comparing their harness running on Astra xhigh to Codex with Astra max?
---
Also worth noting that OpenAI just added support for async tool calling to their harness, which isn't 1:1 with this approach, but is slowly ramping up in being able to provide something similar.
A big part of why Codex uses so many tokens is that it basically hot loops on polling tasks it starts for... absolutely no good reason: https://www.reddit.com/r/codex/comments/1wdlp7q/weve_discove...
I fixed it on my fork of Codex too, also back in Jan/Feb – I keep this patch rebased, for anyone who wants it: https://github.com/tekacs/codex/commit/9ffcf8db9078eae43d411...
It results in token savings similar in scale to those displayed here by Unreal.
---
My harness has used a slightly fancier version of the approach that Unreal is using since ~Feb, and... it definitely works excellently, but it's also assuredly smoother with Astra and other recent models that are more aware of async tool calling.
each individual benchmark that is combined into agentic coding index was compared on xhigh between ua / codex / pi, and headline improvement was calculated on xhigh, but then agentic coding index pareto chart by default includex codex max, hense the confusion.
This is available in DSH via PTC mode.
Basically, instead of the completion API returning a sequence of tool calls, it returns a program that invokes the tools.
I think most are returning Typescript program
With ordinary tool calling, you may have to wait for a tool to return its result to the completion API before the model can continue and issue the next tool call.
With PTC, the generated program can invoke independent tools asynchronously, collect their results, and dispatch the relevant results back to the completion API in a single turn.
Claude and Codex added this as well, but there are some papers on this that show it's not good across all tasks, and models need to be trained for this specific technique, which is what AI labs are doing now.
This plainly seems like a trademark issue in progress considering it's in the same exact domain and considering how many others were confused the way I probably was.
People, I'm begging you, please talk to a lawyer before launching.
>Computer software, namely, game engine software for video game development and operation; Computer software, namely, software development tools for the creation of computer-generated imagery and graphics for the production of video games; Computer software, namely, software development tools for the creation of computer-generated imagery and graphics for the production of content for virtual worlds and 3D platforms; Computer software, namely, software development tools for the creation of computer-generated imagery and graphics for the production of motion pictures, television shows, videos, 3D animations, 3D simulations, 3D visualizations, virtual reality motion pictures, virtual reality television shows; Computer software, namely, software development tools for the creation of computer-generated imagery and graphics for the production of virtual reality video games; Virtual reality game software; Virtual reality software for creating multimedia content; Augmented reality game software; Augmented reality software for use in mobile devices for integrating electronic data with real world environments for the purposes of entertainment
Sure you can, trademarks are contextual. If they were a landscaping business it wouldn't matter. But within the same industry absolutely.
How likely is it that they get sued into the ground in a year? They might have a strong suite of offerings even as soon as six months from now, but if the essence of a company's brand seems at jeopardy from the start, can I take the risk as a potential customer that they'd survive that kind of action?
Would be interesting to compare to a harness optimizing for cost reduction too.
What a good time for harness design. Just while OpenCode is growing up a bit and focusing on their harness. I’m delighted to see people focusing on good general solid harness design principles.
Someone make an OpenCode API (the best general agent end UI API I know) compatible server for it!
Sucks that I haven’t upgraded my client to V2 yet and I don’t really want to build on an outdated API ...
Edit: It would be cool to get streaming responses :) But I understand (and actually applaud) that the authors seem to have been very focused on the core mechanics.
Also note that this works with the responses API!
i think the language primitives map well to the natural desire to use these tools across networks and in concurrent workloads.
another Go harness i keep up to date with is: https://github.com/boldsoftware/shelley (exe.dev team)
It's both expensive to test alternative harnesses and it's expensive to develop them (if using API pricing).
It says that it removes tokens wasted while a model is waiting on synchronous tool calls. What tokens exactly is Pi using when waiting?
We found this massively outperforms any kind of agent like this and the official unreal MCP systems. Its similar to the Blender MCP which also just exposes scripting + very minimal api to claude code/others.
Actors can have identities and roles for RBAC, etc. you - cross agent communication is the same as sending any other message to a actors.
Not to mention that actors can be on your device, another device, etc. whatever the router can resolve - it's transparent to the agents.
Let a model get notifications when something it has accessed/ subscribed to changes, a tool produces output or it gets some other kind of message (for example by another AI or flesh agent). Wait until the next turn or wake it up. It can still decide to do nothing and wait.
1. Agents usually depend on output of the commands they're running in order to make decisions about what to do next, so how do they behave while they're "waiting around" for the output they need?
2. Agents can _already_ run software async, via multiple mechanisms: raw CLI tools like "nohup", literally running tools in parallel (I see Sol do this often in the Opencode TUI harness), and using parallel sub-agents to e.g. research in parallel.
Thus I wonder, how much does this really improve speed vs only improving the "appearance" of getting more done faster?