We built this so we could go.
Most agents demand a babysitter. You start one and you stay — eyes on the chat, thumb on the spacebar, lunch at the desk. Syntaxis is the opposite shape: set the rules once, claim an issue, and walk away. By the time you come back from coffee the patch is open, the tests are green, and a single approval is queued. You did the work. You also went outside. desktop beta
you
syntaxis
The first agent was good. It just wouldn't let us leave.
Five of us were in a Lower East Side coffee shop in early 2025, trying very hard not to look at our phones. We'd shipped agentic systems before — they worked, mostly, on a good day, with a fair wind. The catch was always the same: you couldn't go anywhere. The agent would stall on an ambiguous instruction, or try to push something it shouldn't, or quietly do nothing for nine minutes, and the only way to know was to be watching.¹
That morning the espresso went cold three times. Each time someone had to lean back over, read a paragraph, tap a button. The agent worked; the human didn't get a break. We'd traded a junior engineer for a needier one.
What if the agent's whole job was to stop in the right places?
That's the only idea behind Syntaxis. Not a smarter model. Not a bigger context window. Just a careful junior who has read the rules — yours, ours, the repo's² — and who treats every side-effect as a place to stop and ask. Six gates between intent and your main branch. The agent does the part that's mechanical. You stay outside.
We get the same question from every demo: so does it actually let me leave? The honest answer is yes, with conditions. The first run on a new repo wants a witness. By the third issue, you're at the park. By the tenth, you've forgotten how loud the fans on your laptop used to be.³
Tenets we wrote down before the first line of code.
Side-effects always stop at a door.
If the rule isn't written down, the agent doesn't know it.
Read-only is the default, not the safe-mode.
One eligible item at a time. Synchronous by design.
A good log is a good interface. Show everything the agent saw.
Your branch name, your rules. The agent inherits — it doesn't impose.
Calm interfaces. No jingles, no streaks, no urgency.
We dogfood our own gates. Every day has a stop.
// Two on-call gates per week. No Slack on weekends. We run the same six-gate loop we sell — every PR from a teammate goes through the same approval flow as a PR from Syntaxis. The only difference is the prompt.
What we imported from people who got there first.
We hire the way the agent works: slowly, through gates.
Each role takes about three weeks end-to-end. There's a written take-home, a paired debugging session, a manifesto challenge ("which tenet do you disagree with most, and why?"), and a reference call. No leetcode. No surprise live-coding.
rust + ts · 5+ yrs
comp · 145–180k + 0.3–0.6%
linux internals · seccomp · firejail
comp · 160–195k + 0.2–0.5%
ts + node + editor protocols
comp · 130–165k + 0.2–0.4%
we reply within 7 days
The terminal is on. Say something.
info@etenos.ai // gpg key on the keyserver, fingerprint 3f a3 · 7b 41 · 9e 02
careers@etenos.ai // or apply via /05 above