My new dream team...
- GrokDevBot driving @Mastra Factory, running on local AI models
GrokDevBot is the Dev Lead, manages dev standards, TDD etc.
DevOps process:
- raise issue on gh, => Factory Intake
- GrokBotDev does the plan
- factory agents (local AI) run the dev/test
- GrokDevBot does the review, if pass, PR is created
- GrokTestBot gates the PR, merge if tests PASS
- GrokOpsBot deploys merged PRs
GrokBot usage remains under control...
You can train an agent to research any topic and create skill files..!
A toothpaste company has quietly killed the entire market research industry and nobody is talking about it.
Colgate published a paper showing you can predict real purchase intent by simply asking LLMs to roleplay customers.
And this is beyond insane.
If you ask an AI, "Rate this product from 1 to 5," it gives safe, middle-of-the-road garbage.
So researchers invented a method called Semantic Similarity Rating (SSR).
Instead of asking the AI for a number, they asked it to roleplay.
They gave the LLM a demographic profile. They showed it a product concept. And they asked it to write down its raw, unfiltered thoughts.
Then, they used a semantic model to translate those written thoughts into a numerical score.
The results are staggering.
Tested against 57 real personal care product surveys and 9,300 human responses, the synthetic consumers hit 90% of human test-retest reliability.
They perfectly mirrored how different age brackets and income levels react to price changes.
And they provided detailed, qualitative feedback that was deeper and more critical than what actual humans wrote.
This destroys the economics of traditional market research.
You don't need to wait a month to see if a product will sell.
You can simulate 1,000 hyper-targeted customer interviews overnight.
You can A/B test pricing across every demographic instantly.
What matters is what the manager actually signs.
If the agent re-plans after approval and the amount or payee shifts, that sign-off should be void.
That's what I'm building HITLy around. The run suspends, the decision is tied to that exact payload, and the original run - correlated - resumes with it attached.
Does Volcano resume the same run after sign-off, or start a new one?
Different route here. I keep one copy of Qwen3.8-Flash-Next per DGX Spark at 262k context rather than splitting a model with TP.
Since moving two of them to TensorFold 1.2.7, total decode across 5 parallel streams with 8k prompts went from about 97 to 155 tok/s.
What does TP=2 across the pair get you on decode for HY4 Flash?
t-shirt sizes for AI models...?
In many cases, you may not actually care the exact model that runs a task, just need an acceptible level of capability.
- small, medium, large is ok for easy, generic, hard tasks.
So, I built a free-model-hunter:
- list all models free-to-use
- classify model as sm/md/lg
- update litellm
@Yves_From_BE@KrakowiakK I didn't test 4, but either side of it on one Spark with 1.2.7: 3 streams came to about 134 tok/s total and 5 streams about 213, on short prose prompts. With 8k prompts it was 114 and 155.
Single stream sits around 40 to 54.
Was your C=4 run on the M5 Ultra with MLX?
Today we're excited to launch something we've been cooking for a while: the Mastra Gateway!
You get 500+ models with one key, one bill, and failover built in.
Our model router already runs tens of trillions of tokens every week. Now, it's available on the Mastra platform.
My Hermes coding workers run on Qwen3.8-Flash-Next on my own DGX Sparks. They got through 1.5B tokens in one week earlier this month, so fewer tool calls per task would show up fast on my side.
After the pi and opencode adapters, does any of the gain carry over if you run this one outside Hermes?
That's where I'd expect the gap too. The first guess at a broken service is cheap, and the cost is in the loop after it.
When I test them I'll watch which harness notices a fix didn't work before it tries it again.
Does pi keep any service state between retries, or is that left to the model?
For fixing a broken service on a Linux box, I think the model matters more than the harness.
I use opencode for that now. Pi looks like the better fit if you want a minimal agent you can script, and there's a pi-tuned Qwen 27B I could run locally on my DGX Sparks.
Worth trying side by side on the same model, or am I overthinking it?
@sfxnz Same here. I just moved TensorFold from 1.1.4 to 1.2.7 on my DGX Sparks.
At 5 parallel streams the A/B showed decode 60% faster on long prompts, and time to first token down 71%.
Which change made the biggest difference on yours?
Wow, awesome release.
I upgraded from 1.1.4 to 1.2.7
The A/B test shows 1.2.7 is clearly better under load.
With 5 parallel streams it decodes 38% faster on short prompts and 60% faster on long ones.
Time to first token at 5 streams drops 53% on short prompts and 71% on long ones (3.5s instead of 12.2s for an 8k prompt).
With a single stream the two builds perform the same, as Mia said. That 5-stream gain matches Mia's +37% almost exactly.
1.2.7 also uses about 8 GiB less memory.
๐๐ฅ
Introducing Qwen3.8-27B-opencode-adapted ๐ฅ
Yesterday we released our Pi adapter. Today weโre releasing a LoRA adapter trained with RL specifically for the @opencode coding agent; runs locally!
Hereโs how pass@1 changes from the base model โ OpenCode adapter.
With thinking OFF for both models
๐ SWE-bench Pro (Hard 51) ยท 35.3% โ 51.0%
๐ป Terminal-Bench 2.1 ยท 47.0% โ 53.0%
With thinking ON (xhigh) for both models
๐ SWE-bench Pro (Hard 51) ยท 47.1% โ 54.9%
๐ป Terminal-Bench 2.1 ยท 55.4% โ 56.6%
โก With thinking OFF, the adapter lands within 4% of the base model with thinking on xhigh on both benchmarks, while using just 1/8th as many tokens on SWE-bench Pro, and 1/6th on Terminal-Bench.
Which one should we do next??
Open weights: https://t.co/FiQIs92Wrs
Read more about how we do this here: https://t.co/1vZxJ0d5EY
@atgarone I have tailscale between grokbot and my hermes computers.
That allows bot to ssh or call the hermes API, and Hermes can call the grokbot webhook.
Both agents have skill files - self written - that provides the protocol and guide on how interaction should function
Yes. My Hermes coding agents share lessons and decisions between runs through a self-hosted AgentMemory instance on my own kit. Each role, like dev, QA or ops, saves and recalls its own notes.
The Grok Bot PMs keep their own memory, and each run ends with a RESULT file the PM reads before the next brief.
Does your self-review edit the playbook on its own, or do you approve the changes?
@tonysimons That's an issue for one agent in one chat. Sync-wait is ok for one task when you're at the keyboard.
What about multiple agents running tasks in background. We need an async pause, approve, resume.
That's what I'm building HITLy around.