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.
Your point is valid for individual developers or small teams of technical folks.
However, MCPs are a godsend to anyone trying to manage access to systems (especially SaaS systems) in a large organization that includes a ton of non-technical employees. We use at least a couple of SaaS products where RBAC is severely limited in what it can govern or where API keys do not have any kind of governance so handing one out is basically handing out read/write on the entire system. Being able to configure an MCP to only have read endpoints enabled it a very useful piece of duct tape. Obviously it would be great if the underlying API was just designed with proper governance but that's out of the user's control.
dvt · · focus · HN ↗
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.
ravenstine · · focus · HN ↗
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.
rockostrich · · focus · HN ↗
However, MCPs are a godsend to anyone trying to manage access to systems (especially SaaS systems) in a large organization that includes a ton of non-technical employees. We use at least a couple of SaaS products where RBAC is severely limited in what it can govern or where API keys do not have any kind of governance so handing one out is basically handing out read/write on the entire system. Being able to configure an MCP to only have read endpoints enabled it a very useful piece of duct tape. Obviously it would be great if the underlying API was just designed with proper governance but that's out of the user's control.