Log In
Sign Up
# Agentic AI Engineer
## AI Labs
“`
location: 17.3850° N, 78.4867° E
yoe: [1, 3] // inclusive on both ends, yes we mean it
mode: on-site // we’ll get to why
status: open until we find 3 of you
“`
> if you had to look up the coordinates, that’s fine. if you had to look up why we wrote the role this way, this probably isn’t the role for you.
—
## the work
We’re building agents that do actual work. Not chat features. Not “AI-powered” anything. Agents with budgets, latency targets, and on-call rotations for when they break.
If you’ve ever stared at a `tool_use` block that hallucinated a function name and felt that specific kind of shame, you already know the job.
AI Labs is a small, deliberate team. The charter is broad on purpose: coding agents, copilots, workflow automation, retrieval-heavy systems, multi-agent orchestration, and the eval harnesses that catch a frontier model release silently regressing your production loop on a Tuesday.
The “will agents reshape software” debate is over. We’ve decided. We need people who can ship.
—
## the first 90 days, probably
You’ll touch most of these. Some go to prod. Some get killed at week 2. That’s the job.
* take an agent end to end. problem framing, tool design, eval, rollout, and actually measuring whether it moved a metric. not a demo. a loop that runs while you sleep.
* build the orchestration layer that routes work across `Codex` for codegen, Claude for the hard reasoning, a small model for high-volume classification. with explicit budget, latency, and accuracy SLAs per task. cost matters. latency matters. we track both.
* stand up evals that catch regressions when you swap a model, rewrite a prompt, or rewire a tool chain. not a notebook. a pipeline. with assertions, not vibes.
* replace a process. pick a manual loop (ops, support triage, code review, anything). end it. we’ll measure the hours saved and put it on a leaderboard.
* make a stupid problem and a smart model useful to each other, which is honestly most of this job.
Beyond that, you tell us what to build. The roadmap is not a menu we hand you.
—
## must-haves (we mean it)
* **1 to 3 years of professional experience.** not interns. not seniors pivoting in. the role is calibrated for early-career builders already shipping.
* **you live in or are moving to `17.3850° N, 78.4867° E`.** on-site. we don’t believe great agentic systems get built in async slack threads, and we’re not pretending otherwise.
* **hands-on, opinionated experience with all of these:** `Codex`, `Cursor`, `Claude Code`, `Claude Cowork`, `Antigravity`. hands-on means you have a take on which one wins for which workflow. not “i watched a YouTube demo”.
* **multi-LLM fluency.** you’ve shipped against at least three of: Claude (Sonnet, Opus, Haiku), GPT (4o, 4.1, the o-series), Gemini, DeepSeek, Llama, Mistral. you know which one to reach for to extract structured data out of a soup of HTML, and which to reach for to write a 2000-word doc. and you know why.
* **real agentic work.** you’ve built something that plans, calls tools, recovers from failure, and produces an outcome. you can describe the loop architecture, the eval, and what broke first. (it’s always the JSON parsing. it’s always the JSON parsing.)
—
## things you’ll nod at if you’re our person
read these. if four or more land for you, please apply.
1. you’ve watched a context window slowly collapse under its own weight at step 30 of a 50-step plan
2. you’ve debugged a loop that wouldn’t terminate and you knew, in your soul, that it was your fault and not the model’s
3. you have an opinion on `temperature=0` that takes more than one sentence to explain
4. you’ve felt the specific 3am dread of an agent burning through your token budget on a task it can’t complete
5. you’ve shipped a prompt that worked perfectly on Friday and was quietly broken by a provider update on Tuesday
6. you know the difference between an eval and a vibe check. you also know that most teams ship the second.
7. you’ve reached for MCP and then reached past it
8. you’ve put two agents in a room together, regretted it, and then done it again because it was the right call
9. you read provider changelogs the way other people read sports headlines
10. you’ve felt smug for ten minutes after replacing a 400-line prompt with a 40-line one that scored higher
—
## bonus signals
* you’ve shipped a coding agent or dev productivity tool. even small. even just for you.
* your last role was at a startup or product company. not services. not “we did some PoCs for a client”.
* a public footprint: GitHub, blog, X, a discord where you’re known. we will read it.
—
## you’re probably not our person if
* your AI experience is “i added an OpenAI call to a Django app”
* you’re an ML engineer whose last three years were training pipelines and feature stores. real respect, but this isn’t that role.
* you list one provider on your CV. that signals you haven’t compared.
* you want a research role with whitepapers as the output. we ship to prod. every week.
* you need a spec’d ticket to start working.
—
## how to apply
skip the cover letter. send us this instead.
1. **one agent you’ve built.** a repo, a Loom, a deployed URL, anything. if IP stops you sharing, write 300 words on the most interesting agent design decision you’ve made and what you’d change next time.
2. **one opinion.** pick any two from `Codex`, `Cursor`, `Claude Code`, `Claude Cowork`, `Antigravity`. in three paragraphs, tell us which one you’d build a $10M production system on, and why. (this question is partly a rorschach. that’s fine.)
3. your resume. separately.
“`
to: careers@grabon.in
subject: tool_use: { “name”: “apply”, “input”: { “candidate”: “[Your Name]” } }
“`
we read every application. interesting work gets a reply within 5 business days. uninteresting work gets a reply too.
we know some of you will use an LLM to help write this. we’d be hypocrites to mind. just don’t let it sand off your voice. we can tell.
—
“`
# TODO: take this posting down once we’ve hired 3.
# if you can still see this, the seat is still open.
“`
