Vladimir Kosenko · Octomind (@web3brightside) header
Vladimir Kosenko · Octomind (@web3brightside) profile picture

Co-founder @Muvonteam. Building Octomind, the open-source agent runtime for work that keeps going after you close the laptop

Joined December 2024
Photo by Vladimir Kosenko · Octomind (@web3brightside)Photo by Vladimir Kosenko · Octomind (@web3brightside)Photo by Vladimir Kosenko · Octomind (@web3brightside)Photo by Vladimir Kosenko · Octomind (@web3brightside)Photo by Vladimir Kosenko · Octomind (@web3brightside)Photo by Vladimir Kosenko · Octomind (@web3brightside)Photo by Vladimir Kosenko · Octomind (@web3brightside)Photo by Vladimir Kosenko · Octomind (@web3brightside)Photo by Vladimir Kosenko · Octomind (@web3brightside)Photo by Vladimir Kosenko · Octomind (@web3brightside)
I'm Vladimir, I build Octomind at Muvon: you hand an AI agent a job in one sentence, the result lands in your chat. This one watches three pricing pages every Monday and reports only what changed. 33 cents a run. Every post here is a real job and its cost.
2
1
1
603
4 lines i add when an agent runs unattended 1. do not contact anyone 2. never post anything, send me the drafts 3. write it to report.md, give me the short version here 4. name the field. "at least 500 km" on a car search got read as range, "mileage at least 500 km" didn't
1
16
1 to 3 ran in october on 20 companies (researched, nobody contacted), 5 post drafts (none posted) and a stripe vs paddle report that landed as a file plus 3 sentences in chat. the 500 km was meant to skip brand new dealer cars
11
didn’t know it had snowed until i stepped outside and got a face full of wind-blown snow. five more months of this
19
united airlines now bans third-party ai agents from booking flights for you. its reasons: a shared login can expose your data, and an agent could buy things or spend your miles by mistake. i'd check a site's rules before pointing an agent at it
1
28
ran octomind and codex cli on the same model and the same real-PR tasks. both solved 38 of 65, octomind used a quarter fewer tokens
16
38 remote roles that mention rust or AI agents. an agent pulled them off HN's jobs page and the who is hiring thread, with links. not job hunting, i wanted a list of who's hiring for this kind of work. it saved the list and now only pings me when a new one shows up
1
43
the prompt was "list only remote roles that mention Rust or AI agents, one line each with the link. save what you found to seen.md. this is the baseline, on later checks message me only about new ones". $1.12 of tokens for the first pass
14
octocode, the code map i built for agents, never sends your code anywhere, and i made the worse-scoring local model the default. 0.835 on our retrieval test against 0.992 for the cloud stack. 518 stars so far, i think mostly from people whose code can't leave the building
2
1
52
local embeddings (fastembed), no API key, no reranker. our own benchmark - Hit@10 0.835 fully local, 0.992 with the cloud stack. a clean clone of 97 files indexed in 319 seconds
20
Vladimir Kosenko · Octomind retweeted
One of our agents asked 101 times in a single turn whether the same file existed. Since Octomind 0.61, the supervisor ends the turn after 10 consecutive rounds of identical calls with identical results. muvon.io/blog/release-round-…
1
2
21
down in the safeguards part of the fable 5.2 announcement - when a cyber question gets blocked, "users will instead receive a response from Opus 4.8". blocked biology ones go to opus 5. i read the paragraph twice looking for the bit where you're told the model switched
51
fable 5.2 is out, 20% cheaper than 5.1 at $8 in and $40 out per million tokens and the launch post itself says to start with opus 5.5 for most coding, then move up when the task needs it. i'd do that, fable for the jobs opus gives up on
1
195
Codex lost versus Octomind All the results are open source
17
Goldman took Claude Code away from most of its traders and bankers in October, engineers keep it. Bloomberg reported it plus in July their CIO was praising the non-technical staff for vibe coding their own tools
2
28
the in-house assistant is still there for the whole firm. what got restricted is the outside agents that write and run code on their own
15
Already using it and isn’t really cool for writing posts
AI: Now I understand your thoughts right ... In 5 mins same generic opener and same shit post So we build Unsame that helps your AI brother never make the post the same unsame.io
19
Seems like X is X (dead, not loading)
1
10
asked an agent to check 12 AI answer tools without a login. 70 cents and it came back with all 12 blocked and the list. the part that made it say so was one line in the prompt - "say which tools you reached"
2
23
perplexity, genspark and chatgpt search gave it 403s, you.com wanted a login, 6 were js shells with no answer text, phind 403, lepton unreachable. it said so, then labelled a plain web search as the fallback instead of passing it off as the real thing
99