Black Friday deposit bonus·Top up $50+ and get $10 · Top up $100+ and get $25·Ends November 30Top up →

Let an AI agent fill a job application for you

Use a browser agent to complete multi-step job forms, upload your CV, and handle screening questions on real sites.

Nov 11, 2026 · 8 min read

A custom corporate ATS will not give you an API. Use an autonomous browser agent instead. It reads the DOM, selects radio buttons, types from your profile, and attaches your PDF resume, and submits the form the way a person would.

Pair vision-capable models with a headless browser runner. You can combine Claude, GPT and Gemini on one prepaid API key through the browser-agent interface. Lovely Router (lovelyrouter.com) puts web data tools, a browser agent, video models, and Claude, GPT, and Gemini for AI agents on that same prepaid key, so the form runner, the answer writer, and any company research share one balance and one endpoint.

Career sites are built for people, not for scrapers. Hosted applicant-tracking portals and in-house ATS pages hide fields behind custom components, lazy-loaded steps, and session cookies. An API-style job board client fails the moment the markup changes. A browser agent does not need the vendor's schema. It opens the public apply URL, waits for the page to settle, and interacts with the same controls a candidate would click.

1. Configure the browser agent for form navigation#

The agent needs a starting URL, your profile data, and the path to your resume. Give it clear instructions for form fields, dropdown selectors, and multi-page pagination.

It loops through the DOM, matches inputs by label or aria attributes, and types your details. For a file upload, the runner targets a standard <input type="file"> and passes your local PDF path.

Keep the profile structured. Store legal name, preferred name, email, phone, location, work authorization, earliest start date, salary expectations, and links to a portfolio or LinkedIn profile as plain fields the agent can map. Write a short work-history block and a skills list in the same payload so screening answers stay consistent across applications. Tell the agent which fields it may invent nothing for: never guess a visa status, a salary number you did not approve, or a degree you do not hold. When a required dropdown has no exact match, the instruction should be to pick the closest honest option and record the choice in the run log.

Pagination and conditional fields are the usual failure points. Some forms reveal the next section only after a radio button is set. Some disable the continue button until every required asterisk is filled. Spell those rules in the task prompt: fill visible required fields first, scroll the active step into view, click continue only when the button is enabled, and stop if a validation message appears. For selects that are not native <select> elements, instruct the agent to open the list, type to filter, and click the matching option rather than setting a hidden value.

This Python snippet uses requests to start the session:

Python
import requests

url = "https://api.lovelyrouter.com/v1/browser/run"
headers = {
    "Authorization": "Bearer YOUR_PREPAID_KEY",
    "Content-Type": "application/json"
}
payload = {
    "target_url": "https://example.com/careers/apply/123",
    "actions": [
        {"type": "navigate", "url": "https://example.com/careers/apply/123"},
        {"type": "fill", "selector": "input[name='full_name']", "value": "Jane Doe"},
        {"type": "fill", "selector": "input[name='email']", "value": "jane@example.com"},
        {"type": "upload", "selector": "input[type='file']", "file_path": "/path/to/cv.pdf"},
        {"type": "click", "selector": "button#next-step"}
    ]
}

response = requests.post(url, json=payload, headers=headers)
print(response.json())

Treat that payload as the deterministic core. Navigation, known selectors, and the file upload stay scripted so the same resume lands in the same control every time. Leave the open-ended questions to the model step described below. If a site uses a shadow DOM or an iframe for the application widget, note the frame in the task so the runner looks inside it instead of the parent page. After the run, store the final URL and any confirmation text. A later script can open that receipt if you need to prove the application was submitted.

Model choice affects both quality and cost. Claude Sonnet 5.5 (claude-sonnet-5.5) is $2.00 per 1M input tokens and $10.00 per 1M output tokens, with cached input at $0.10 per 1M. Claude Haiku 5.5 (claude-haiku-5.5) is $0.10 per 1M input and $0.50 per 1M output when the page is simple and the answers are short. GPT-6.1 Sol (gpt-6.1-sol) matches the $2.00 / $10.00 band, and Gemini 3 Flash (gemini-3-flash) is $0.50 per 1M input and $3.00 per 1M output if you want a cheaper pass over long DOM dumps. The full list lives on the models page, with family pages for Anthropic, OpenAI, and Google.

2. Handle multi-step flows and custom screening questions#

Job applications rarely fit on one page. Most career sites split the flow into contact information, work history, resume parsing verification, and custom screening questions.

When the agent hits a question such as "Describe your experience with distributed systems," it calls an LLM to draft a concise, relevant answer from your uploaded CV text. It types that answer into the text area, checks the character count limits, and moves to the next page.

Prompt the writer with the job description, the exact question, the character or word cap, and only the resume sections that apply. Ask for a first-person answer that cites tools and outcomes already on the CV, and forbid new employers, dates, or metrics. If the box allows 500 characters, tell the model to stay under that cap so the agent does not have to truncate mid-sentence. For yes/no legal questions, do not let the model improvise: pass a fixed boolean from the profile and click the matching radio.

Resume parsers on the ATS sometimes misread the PDF and pre-fill wrong dates. Add a verification step. After upload, read the parsed fields back from the page and compare them to the profile. Correct mismatches before continuing. That extra read is cheap next to a rejected application, and it uses the same page-reading path as the rest of the run.

If login requires 2FA or a CAPTCHA, the agent pauses, captures a screenshot, and surfaces the challenge in your control center. You complete that step by hand, then hand control back to the run. Do not ask the agent to defeat CAPTCHAs or to reuse someone else's session. Create the account yourself, finish the human check, and resume. The same pause pattern works for "are you a robot" gates and for sites that email a one-time code.

Some employers add voluntary demographic questions. Leave those blank unless you have an explicit preference stored in the profile. Equal-opportunity forms are optional on most sites, and an agent should not invent answers. For work samples, only upload files you named in the task. A portfolio link in a text field is safer than attaching the wrong PDF.

You can research the role before the form opens. Web search on Lovely Router is $0.00090 per search, and reading a page is $0.00090. Pulling a job description, a careers FAQ, or a public engineering blog gives the screening writer better context without a separate vendor key. Those tools are listed on the tools page, and the pattern fits web automation as well as recruiting data when you are collecting openings rather than applying to them.

3. Calculate the run time and token costs#

A full application run covers DOM analysis, vision processing to locate elements, and text generation for screening questions.

To price one run, tally the input tokens used to read the page DOM, the output tokens for action plans and written answers, and any web search used to verify company details.

ItemUsageUnit PriceCost
Agent Input Tokens120,000 tokens$2.00 per 1M$0.24
Agent Output Tokens15,000 tokens$10.00 per 1M$0.15
Page Fetch Operations8 pages$0.00090 per page$0.0072
Total$0.3972

A standard run takes about 45 seconds across three pages and costs under forty cents in combined model and tool fees. Parallelize your runner scripts to cover dozens of target roles.

The table assumes a mid-tier model at $2.00 per 1M input tokens and $10.00 per 1M output tokens, which matches Claude Sonnet 5.5, Claude Sonnet 5, GPT-6.1 Sol, and GPT-6 Sol. One hundred twenty thousand input tokens are 0.12 million, so 0.12 × $2.00 = $0.24. Fifteen thousand output tokens are 0.015 million, so 0.015 × $10.00 = $0.15. Eight page reads at $0.00090 are $0.0072. Adding those lines gives $0.3972 for the example.

Cache repeated profile text when the provider supports it. Claude Sonnet 5.5 cached input is $0.10 per 1M tokens, one-twentieth of the uncached input price. If the same resume and instructions are sent on every application, cache reads shrink the input line without changing the answers. Claude Opus 5.5 is $4.00 per 1M input and $20.00 per 1M output if a hard form needs a stronger planner, and Claude Haiku 5.5 stays near a few cents when you only need short field fills. Use the calculator to swap models before you batch a week of roles, and read Claude prompt caching before you design the repeated prefix.

Add tool calls only when they change the application. One web search at $0.00090 and one page read at $0.00090 barely move the total. A screenshot of the confirmation page is $0.0027. Ten applications at the table's rate are 10 × $0.3972 = $3.972 in model and fetch fees, before any extra research. That is the unit cost to plan around, not a quote for every site. A multi-step corporate career-site flow with more steps and longer screening text will use more output tokens. A one-page form will use fewer.

Watch rate limits when you parallelize. Space runs so you do not open dozens of sessions against one employer domain at once. The guide on Claude API 429 errors covers backoff if the model side returns too many requests. Keep one prepaid key, log token usage per application id, and stop a batch if confirmation text is missing. The agents docs describe how to attach tools and models to a longer-lived worker if you outgrow a single HTTP call.

Next steps#

Use browser automation in your recruiting workflows, or test your own multi-page web forms against autonomous agents. Fund your account with a single payment and access all models and tools from one endpoint on the pricing page. Start with one role you already intend to apply for, confirm the uploaded CV and the screening text by hand, then widen the list. Integration notes are on the docs and integrations pages if you want the same runner inside an existing script or editor.

Try it yourself

One prepaid key, no subscription. Every price in this article is the price you pay.

Keep reading

All articles →