I'm starting a business with a team of Grok Bots.
The app helps founders get warm intros when they raise. A few founders already want it. Now it needs to become a real product, and I'm not hiring yet.
So the first hires are bots.
2
1
49
Replying to @Tech2Wild
What are you thinking? I’m on the same page as you. Curious about how you like the micro duck, I’m sure we’ll get a video!
21
Replying to @kyzoroX
This is the number that matters for agents. Every turn resends the tool schema and history, so you pay that prefill over and over, not once. tok/s is the wrong thing to shop on
2
Replying to @MiaAI_lab
20GB of weights with 3.4B active is the part I care about. Leaves real room on a 32GB card for the long context agents need. My 5090 shows up today, this is going on it first
55
Replying to @dantelex
Makes sense, planning is where you want it slow and picky. Do you hand the plan over as a file or just paste it in?
4
The expensive part of a bot team wasn't the models. It was the bots rereading context and talking to each other too much.
A cap per bot, plus a propose-then-approve stage, cut more spend than swapping models.
14
Replying to @0xReso
25 agents on one issue is the part I'd cut. Mine are a few bots with one narrow job each and they still talk to each other more than I'd like. Were the 5 failures the same thing or all different?
1
8
Replying to @MarsHomestead
Two businesses on one bot team is the part I'd want to hear more about. Separate rules for each, or one shared set? Mine each have a written manual only I can edit and they reread it every night
5
Replying to @dantelex
Optimistic vs pessimist is a good way to put it. For execution you want the pessimist, it's the one that checks before calling it done. Ever tried planning in Codex and handing the plan to Claude to build?
1
1
16
Usage reporting zero is the sneaky one. Nothing errors, the budget math is just quietly wrong. Have you found a server that gets all three right?
16
Good, no re-quanting is the big one. Does it serve an OpenAI-compatible endpoint too? That's what decides if it's a drop-in under an agent harness or a rewrite
1
16
Replying to @QuixiAI
25-30 tok/s at 256k on a single 5090 is wild. How do the 3 concurrent requests hold up as context fills? That's the number I care about once a few agents share one box instead of one person chatting.
10
Replying to @dantelex
Product vs harness is the right split. The harness is the part I care about, since that's where subagents, permissions and memory live. When you say design, do you mean UI work or plan quality?
1
25
Interesting timing, the same week llama.cpp hits 1.0 with stable interfaces. What I'd want to know before switching: does it keep GGUF loading and an OpenAI-compatible endpoint, or is it re-quant everything?
1
2
144
Replying to @0xDevShah
Loop vs environment is the right line to draw. What does the env give you that a shared state store plus an orchestrator doesn't? My bot team runs like an org chart (chief of staff + leads, one job each), so coordination is what I design around most.
9
Replying to @penberg
The question I'd add: how many agents at once. One coding agent on a ~27B fits a lot of boxes. Run a few in parallel and memory bandwidth and KV cache start making the decision for you. I keep a list of open models worth downloading, and the sizing column ends up mattering more than benchmarks.
1
1
11
Replying to @shmidtqq
Routing is half of it. The other half is scoping. I split my startup's bots into one narrow job each (product, growth, ops) with a chief of staff on top, so most runs are small by design instead of one generalist re-reading the whole repo.
17
Replying to @JakeSaasing @AlexGDimakis
Honestly I don't fully know yet. Each bot has a written manual for its one job that only I can edit, and they reread it every night and flag work that conflicts with it. Overlap should show up there first. We're still in the propose stage, so I see everything before it ships
3
1
32
Replying to @holonrun
7/8 on tests but timing out at 41 to 72 turns is the stat people should see before picking a model off a leaderboard. Scoping the contract per agent makes sense. Are you splitting skills per agent by hand or loading them on demand?
1
10
Replying to @lankford_gary @alex_verem
16 sec for a 15k prefill on the E4B is solid for a Mini. The tools doubling it tracks with what I've seen, the schema eats context before the model does anything. How many tools are you loading?
18
