Building Electros | a local-first AI agent harness that runs your computer. 2× DGX Sparks + M5 Max Building in public, numbers included
Joined March 2018
- Tweets3.1K
- Following1.3K
- Followers424
- Likes5.3K
totally forgot to pay for parking until i was a good bit away from the car...
left a voice not to my local agent at home running on dual DGX Sparks
"hey figure out how to automatically pay for parking when i leave the house in X carpark"
its been cranking for hours,
At this stake its built a full captcha solver with vision and realtime browser runner, its blasting captchas probably 3 out of 5 times. but its mega persistant
Next stage wire Temp payment card
Hook up into the existing "Ive left the house automation"
when i leave between work hours ill get asked if i want to pay book parking automatically/
... lets se if this ever completes
fun experiment never the less
1
1
51
I think my agent channeled you tweet haf way through its work....
funniest part was when it started reading the text to me
"Chinese character, Chinese character , Chinese character " and so on for 45 seconds
12
here it is, the personal agent
the trend is clear, OpenAI, Meta, X -> all launched their "Always on Agent" in the past month
but it didnt start here
lets run it back - engineers been doing this for a while
- 2025 we transitioned to agent runtimes, claude code on a server wrapped with vibes for remote access
- Jan 2026 - Openclaw launches, full of vibes but mega fun and requires a degree in engineering and patience
- Spring 2026 Hermes, Ironclaw
- Summer 2026 always on agent runtimes but still enginnering grass roots systems
- Fall 2026 consumer grades agents go live
this is is this is the start of AI for everyone, but with a big subscription - its the break out moment for AI and the User
for the engineers we are still building our own custom agents, running on local models, on machines we control with full privacy .
Why is this important
No-one sees my prompts
The UX is mine to own
It connects to what i want
Its got access to what i give it
Its runs on my own hardware and inference,
Local might not be for everyone, but for some of us decentralised personalised, customisable and on-prem and secure is what we build
Love the trend of personalised alway on agents but not paying a subscription for my brain to go to the cloud provider
41
yea totally agree,
acquiring hardware for local AI has been a priority for me since comments like this came out from the labs
lets hope the Open Weights revolution, primarily thanks to China does not slow down, the grass roots optimisation layer is a god send for most of us!
74
I really like it, it's flys along.
The trajectory view is great, basically chrome dev tools for Inference runs.
the way verbose execution his hidden away bit the touch of a button is great.
Taken a lot of inspiration from it for my own harness electros
1
1
45
It's Saturday, my computer is not mine for the day.
After a week of daily driver usages ive identified the next wave of features and bugs to fix for electros harness
Naratted a list of features over a morning coffee, local tts model transcribed, triggered a goal then Stacked up a tonne of features to be built. Electros recursively self improving for the day.
Chaotic to watch.
goals currently being orchestrated running on glm-5.3-flash on dual #DGX sparks using @MiaAI_lab recipe
Dispatching Opus5.5 jobs to Claude code cli, some handled locally some by other frontier models, building fixing prototyping, judging. Shipping
The mobile client pings for approvals if stuck (over tailscale)
What a time to be alive, software building its self locally, farming out to frontier when it feels it's the right time.
This loop.gets better every week
1
118
Electros harness sooon...
- Desktop
- web
- mobile
- browser extension
- cli
Building its own landing page
Using GLM5.3 on DGX sparks
90
zcash:native $1500
what a site to behold. only 8 years waiting for the breakout :-)
Good times - running a jet engine in a unused room in the house circa. 2018
59
mode like opus then fable,
but if you bundle the deepseek harness in, with subagents, plans and keep an eye on it it runs pretty well and its pretty quick.... just does a lot of thinking
Had it running nearly 24 hours on my, AI harness its done about 3B tokens, fixing bugs, adding deatures, full code audits
price goes up tomorrow
1
43
Claude refuses point blank to swap my ETH, ive a backend approval service,
it just wont wont call the MCP tool... point blank NO!
Local models "yes sire, yolo"...!
58
tine to jailbreak our TV's
complete spyware in the house
youtu.be/6IFVTcM28KA?si=FGmo…
1
1
114
deepseek-v4.1-flash-expires-on-0910
super cheap and crazy fast.
running this hot with 100M+ tokens for a few $ is wild.
been claud' maxed out full weekly limits until friday...
2
2
254
GLM 5.3-flash runs executes the goals while i sleep,
it might be slow but it get shit done.
plan with frontier,
execute with Local +Openrouter / minimax for fast subagent with specific specialities
auto pilot,
1
94
