I built an AI agent that applies to jobs for me — tailors my resume per job, fills the forms, and waits for my approval before submitting

I would not let it submit autonomously either, at least not until the system has boring guardrails around side effects.

The useful boundary is: the agent can prepare an application package, but submission is a separate privileged action.

A few things I would add before trusting it with more volume:

keep a per-application receipt: job URL, company, role id, generated resume hash, answers used, timestamp, and whether it was submitted

diff the generated resume/answers against a base profile so weird hallucinated experience is visible before review

make every external write idempotent where possible: do not resubmit the same job/company pair unless explicitly allowed

hard-block fields that should never be invented: citizenship, work authorization, salary, disability/veteran status, location, availability

store screenshots or DOM snapshots of the final review page before submit, because ATS forms are messy and sometimes hide required fields/state

throttle and dedupe aggressively so it does not become accidental spam

The part I would be most suspicious of is not filling text boxes. It is field interpretation: dropdowns like "authorized to work," "sponsorship required," "willing to relocate," or knockout questions where a wrong mapping has real consequences.

So I think your "stops before submitting" choice is the right product shape. The interesting next step is probably a review UI that shows provenance for every filled field: source from my profile, inferred from resume, generated from job post, or needs human input.

the ATS form handling is genuinely the hardest part and most people building these underestimate how bad it gets. Greenhouse alone has like 4 different form rendering patterns depending on how the company configured it. you figured that out the hard way apparently.

the human in the loop before submit is the right call, not just ethically but practically. auto submitted applications read differently, the tone is off, the why do you want to work here answers are always too clean. recruiters are starting to clock it.

one thing i'd push you on though, resume tailoring per job sounds good but are you actually verifying the output before it goes out? because LLMs confidently rewrite bullets in ways that are technically accurate but subtly wrong about your experience. if you're approving the form fill but not re reading the resume each time, that's the gap.

the part i'd think about next is what happens when a company uses a lesser known ATS or a custom built portal. that's probably 30% of roles depending on the industry and your agent just fails silently or skips them. that coverage gap matters more over time as you scale this to more jobs.

curious how you're handling the PDF generation side. a lot of resume to PDF pipelines look fine until a recruiter opens it on a specific version of Acrobat and the formatting breaks.

The stops before submitting part is the important design choice. Job applications are exactly the kind of workflow where agents can save time but the final human review keeps it from turning into spam or misrepresentation.