Josh Stevenson | RecursiveIntell

AI systems engineer building local-first agent runtimes, semantic memory, governed execution & verifiable infrastructure. Founder @ RecursiveIntell.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

3 hermes agents + 10 subagents + 3 chrome tabs running and I still have 5.25GB free. I used to think doing this required 64GB of ram. I have 16 (14 usable). I'm excited for the future after seeing this kind of efficiency. I guess my work on esp32 efficiency shows with this. @Teknium -- I'd be honored if you used my code to lower hermes' RAM & CPU footprint.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

@DanielSmidstrup @AnthropicAI Whichever model you choose as your orchestrator. It even chooses the model/effort in codex/claude based on difficulty of task.

My personal one even adjusts its own skill in middle of night based on the results of the usage that day.

Only possible with the receipts ares creates

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

@DanielSmidstrup @AnthropicAI You can set up hermes/ares to call codex and Claude based on what its doing instead of handling the coding itself if you wanted. All you have to do is spend the time to setup the skills for each and a routing skill so it knows when you expect each to be used.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

@_cherki82_ @f_demaku @thsottiaux https://t.co/nkMyFYC4un
It's a highly custom version of hermes, with the entire backend replaced with my custom stack wrote in rust.

It's a work in progress and has bugs as I'm just one person. Lmk if you have any issues. I run Linux btw.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

@crecross @f_demaku @thsottiaux Ironically, I do fine tune my own models. I however was talking about the harness that controls the model. I wrote the runtime, backend, memory, graph system, specialists and context governor.

This is why I have room to talk. I've done the hard work instead of just crying 😢

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

@f_demaku @thsottiaux Or maybe I'm using the custom harness and agent i created instead of using what someone else made/ spoon fed you and griping you have no real control.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

@f_demaku @thsottiaux I used Astra max + a Sol hard all night and used ~6-7% of my 20x. I have 55% usage left and 2.5 days till reset. I'm gonna have to start using 3-4 at a time to use it all before reset.

Y'all are obviously doing something very very wrong.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

I've been coding with Astra Max all night long other than like 3-4 hours + a Sol Hard session for awhile and I've only used 6-7% usage with Ares (my agent/harness) and the latest updates to my context governor and runtime. I don't understand why others are reporting crazy usage.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

You can actually inspect the PR head on this one. Look at the goal I gave it, then look at the discipline it still imposes on itself while executing.

That’s the runtime doing its job: constraining the agent to evidence-backed actions and refusing unsupported shortcuts, even when taking them would make the goal easier to complete.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

It depends on how it's used. If it just context dumps, then hell yeah. My answer is it being part of the runtime before the model call along with being orchestrated by the context governor and beyond that it being based on a deterministic relevance scoring system before it will even look at the db. All tuned with receipts/provenance instead of guessing on vibes.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

@brockpierson If you didn't have star wars tie fighter or system shock back in the day, we're you even a gamer?

I miss cards like the ati all in wonder card. TV tuner and solid 3d graphics.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

@ChrisTranGG @paulNL That's me bro. I've been building for 9 months what the industry is just now realizing they need. They're all still in python too. My entire stack is rust and I rewrote everything for it. Now I'm creating a custom runtime. Full provenance/receipts. No more black box.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

@stefanopineda @BuckedUnicorn I've heard risers add real latency issues. Maybe more, just something I read a few months back I believe.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

@sickdotdev You get like $1700+ in api token value for that $200 on gpt. I call it the reality of SUBSIDIZED consumer ai.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

@brockpierson My first 3d card was the 3dfx voodoo banshee. It was the first 2d and 3d card. It used the voodoo 2 for the 3d side. Mine was 16mb though

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

@DanielSmidstrup Neither. Moving blindly.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

@eas4ai @Google My memory is fully portable as well. you can fully transfer memory and state to any other agent using the same stack on connected on internet or lan.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

so google is essentially laying out what they see as the future of ai? kinda funny that it is just a crappier version of what I already built with semantic-memory. it's like my draft/prototype build. @google are IP thieves, passing it off as their own. you can trace mine to January/February.

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

I've been coding v2 of my runtime for like a week already and I just iteratively updated the plan.

I have 3 days till my usage reset and I have 62% usage left. So here's hoping Astra Max gets the job done that Luna Max couldn't (lol).

Both are using plans from @openai GPT Pro6

[Josh Stevenson | RecursiveIntell @RecursiveIntell]

@OpenAIDevs @cognition Ares does a lot more than just back up "it works" with tests. It includes full provenance/receipts among other things. Include everything it does as good or better than devin @grok. Also include what it does worse.