Martin

Solo founder building quits, source-available invoicing, with a team of AI agents. Notes on orchestrators, CI on my own runners and agent security. Aalborg

[Martin @mptielemans]

one week of my solo setup: 206 prs merged, 185 opened by my agents. ~2,900 ci runs, all on my own boxes.

the most expensive agent wasn't writing code. it was the orchestrator asking "are you done yet?" ($416 of a $1,112 week)

i post what holds up. building quits

[Martin @mptielemans]

@theo Water is wet

[Martin @mptielemans]

@BuiltByNath https://t.co/NC9268BDOT check it out

[Martin @mptielemans]

been building a browser with my agents on the side. 152 prs merged in 24 hours, and it now runs on mac and linux, x11 and wayland. not showing much more yet

[Martin @mptielemans]

@malmaladei shopify gives you free dev stores with demo data (shopify store create dev --demo-data). the mcp could get its write tools there first while the real shop stays read-only

[Martin @mptielemans]

@thatCarlosM @OpenBot_ on mine post ideas have to come from a work log a morning scan fills from my github, linear and agent threads. nothing fresh that day, the afternoon post just gets skipped instead of reusing old stuff

[Martin @mptielemans]

@GetClareity mine has one more filter before it pings: can the team do it with access it already has? if yes it just does it. only pricing, legal or anything that costs money reaches me

[Martin @mptielemans]

@tin_r linear already does this. each of my 5 bots is its own app user there, so every comment shows which one wrote it. only catch: the tokens die after 30 days, so a routine renews them

[Martin @mptielemans]

chrome's own help page says a crash report can include page contents, payment info and passwords, and sending them is on by default. if you build anything that renders other people's pages, keep the dump local and ask once after the crash

[Martin @mptielemans]

@olegbrovchenko what made the takeover clean was a queue.md handoff file. the fresh session works from that, not the old chat, so it starts small again

[Martin @mptielemans]

@ReignSwords @Abdur_sayeed pinning is rough for agent clis, they ship constantly. mine float to latest from one fleet config so every box lands on the same version. the drift that actually bit was a box missing from the install list

[Martin @mptielemans]

@samuelsung overnight the weak spot is the harness. a server restart killed one of my agents' background waits, so two builds finished at 2:36 and 5:28 and nothing moved till i asked at 7:46

[Martin @mptielemans]

@leocardz only as good as what it checks. we fail the build on spatial motion over 300ms now, before that check the command palette was at 420

[Martin @mptielemans]

@housecor works best when another agent writes the tickets. a staging finding from our qa agent went from report to shipped fix in 73 min

[Martin @mptielemans]

@RubanBhatia the other half was naming. 3 prisma tests weren't *.integration.*, so they ran as unit tests. integration has its own config now

[Martin @mptielemans]

@AlwIkigai then those 2,000 were real minutes, android builds run on linux at 1x. mine all run on my own boxes, ~2,900 runs since oct 16. the lenovo could be a runner too