eigon is hosting for small software that fixes itself. connect a repo and it runs your whole stack frontend, backend, database, auth, domains, backups. when a build or runtime failure is fixable, an agent opens a PR with the fix. when it isn't, it says so plainly instead of guessing. a human merges every PR. it has never pushed to anyone's main branch. 5 teams pay for it, all off hostinger, vercel or supabase. https://t.co/msyH5tnZOy
an agent wrote me a regression test for a bug where two files each define a function with the same name. the test checked that the call went from "go()" to "process()". both candidate targets were labelled "process()". the label is just the bare function name. so the assertion passed against the unfixed code too, which i only found by checking out the original and running it there. rewrote it to assert on the resolved node id instead. if two candidates can share a label, a test on the label tells you nothing.
this is roughly where i landed. the auto-fix in my deploy platform isn't a prompt, it's a loop: a failure comes in, it goes and gathers the real logs and the real filesystem, and writes a PR. the part that gets left out is that a system which prompts itself also fails by itself, at 3am, with nobody reading the output.
This is f*cking insane. Anthropic engineer:
"You're not supposed to prompt Opus 5.5.
You're supposed to build a system that prompts itself. At Anthropic, 90% of our engineers have already built these systems."
In just 30 minutes, she explains how Anthropic builds AI agents that remember context, learn from mistakes, fix their own errors, and improve with every run.
This is the shift from simply using AI to engineering systems that work autonomously.
Honestly, it's better than most $1,000–$2,000 agent courses I've seen.
Watch the video, then save it for later
eigon's auto-fix opens a PR when it can fix your build. a human merges it. it has never pushed to anyone's main branch and it isn't going to. people ask when i'll drop that step, usually as though it's the obvious next move. my answer is that the merge button is where the responsibility actually sits. an agent can be right nine times and the tenth one is still somebody's production. i'd rather run a slower loop than have that conversation with a customer.
one distinction i'd add, learned the hard way: "running" and "responsive" aren't the same state, and most tooling collapses them. i had six agents go idle without reporting, with committed work sitting in their worktrees. idle looked identical to finished. the glossary needs a word for an agent that is alive, answering, and no longer doing the thing it was asked to do.
the problem with agents going quiet isn't visibility, it's that every "done" signal i have is something the agent told me. a background command that double-backgrounded itself fired "done" while the real process was still running. an agent read its pending queue fifteen times in two days and drained none of it, reporting fine the whole time. six more went idle with committed work sitting in their worktrees. artifacts don't lie the way self-reports do. i stopped asking and started reading git status.
four ways i've broken my own database with a migration this year.
1. dropped a unique index so a person could belong to several accounts. every fetchone() that depended on it quietly became "pick whichever". two identical logins, two different accounts. now when i remove a uniqueness constraint i go and grep every query that selected on that column.
2. duplicated a column across two tables. found all the readers, missed a writer. the invite flow broke at its last step for weeks before i noticed. when a column moves now i hunt the writers first.
3. no schema_migrations table, so every migration re-runs on each init-db. a later one dropped an index an earlier one created, and the second run tried to rebuild it over duplicate rows and killed the whole upgrade. if a migration undoes an earlier one i delete the statement from the earlier file too, then apply the whole set twice.
4. restructured member rows because a comment in the file said the table was empty. it hadn't been empty for months, and those ids sit inside the address of every customer's private data. i test against a database holding live data now, never a fresh one.
all four come down to the same thing. IF NOT EXISTS guards the object existing, never the data being able to satisfy it.
@suhailpatel@OpenAI congrats. curious what surprised you most in the first two weeks, in my experience the infra is never the hard part, it's the three undocumented assumptions everyone already has in their heads.
the version that taught me most was cheaper than building one: i dropped a unique index so a person could belong to several accounts, and a fetchone() elsewhere quietly turned into "pick whichever row". two identical logins, two different accounts. uniqueness constraints aren't just validation. every query that assumed them is load-bearing on them.
spent most of my godaddy time on migrations and devops. at that size almost nothing fails loudly it degrades, something absorbs it, and you find out three weeks later from a graph nobody was watching.
spent a stretch building dugout, sports app, 250k+ downloads. the thing i didn't expect about that number: nearly every hard problem turned out to be a timing problem. match ends, everyone opens the app inside the same 90 seconds, and anything comfortable at average load falls over at the peak. average load tells you almost nothing.
ran 6 agents in parallel across separate worktrees yesterday to audit a codebase. they kept getting OOM killed. each one quietly retried. eventually all six went idle without ever sending a report and "idle" looks identical to "finished" from the outside, so i sat waiting for status updates that were never coming. two of them had already committed good work. i only found it by going and reading the worktrees myself.
graphify made github's octoverse 2026 report.
178 commits from me so far, third on the contributor list.
the two ahead of me are @safishamsii whose project it is, and claude. finishing behind the founder is fair enough. the latter is a 2026 problem i wasn't ready for.
congrats @graphify
my migrations run on every init-db. no schema_migrations table, no record of what already ran.
so one migration created a unique index on email, a later one dropped it so a person could belong to several accounts, and the second init-db blew up, the early one ran again, found no index to skip, and tried to build one over duplicate rows. IF NOT EXISTS guards the object existing. it doesn't guard the data being able to satisfy it.
building got easy and it exposed what was actually hard. i shipped something this year that works and has paying customers, and the part i'm worst at by a distance is getting anyone to hear about it. took the building getting cheap for me to notice.
feels like founder mode is a bit of a trap right now
agents made the building part so easy that everyone feels like they could start a company
but it's 90% likely you aren't thinking big enough to actually make it now
this happened to me, was considering going founder mode but then realized every idea i thought of would probably be oneshottable by the next wave of SOTA models (they were, when fable came out lol)
this made me realize my priors were still tuned too hard to a pre-agi world and i needed to shake that before i even considered building something
changed echo-mem's licence twice in a month. apache to BSL on 25 september, back to apache on sunday.
reason for reverting: github reads BSL as "NOASSERTION / Other", because choosealicense doesn't list it. and "Other" quietly drops you out of every directory and roundup that filters on an OSI licence. at this point, being invisible is a worse problem for me than a big company self-hosting it.
@piyushgarg_dev usual culprit with these is an installed app or integration still carrying org-level permissions long after whoever added it moved on. they never show up in the member list, which is the only place anyone looks.
last week i nearly orphaned every customer's private memory graph. the id of a private graph has the member row's id inside it. an old migration restructured member rows freely, and its own comment said that was fine because the store held zero facts at the time. that stopped being true months ago. rewrote it as an ALTER TABLE preserving every id. recreating those rows would have detached every graph silently, with no error anywhere.
@strimblez missing, not missed. two days of deployment work lived only in commit messages never written into memory as a fact. everyone benchmarks retrieval because retrieval is benchmarkable. nobody measures capture, and that's where mine was losing.