2026 · Agentic AI and service design · 5MIN READ
Job Pipeline: an agentic job search I designed
I designed an agentic system that finds, checks and prepares job applications for me, so I can spend my limited hours on the decisions only I can make.
3 agents, 1 tracker, 1 human decision point
UX Job Digest
Scheduled
Searches job boards and job-alert emails, checks every role on the employer's own site, scores the fit.
Job Inbox Check
Scheduled
Reads my inbox, updates application stages, flags who is worth a follow-up.
Prepare Applications
On demand
One button writes a tailored CV, CV summary and cover letter, then audits them.
Tracker
Web app with its own database
Every role, score, kit and stage in one place.
Me
Decide, edit, send, close. Nothing leaves without me.
The problem
A job search is a full-time process. My working window is Kita hours, 8:30 to 16:00, shared between an AI design course, launching my own business and the search itself. In my manual search between mid-August and late September 2026 I saw four problems:
- Scattered sources. Roles were spread across StepStone, Indeed, Glassdoor, the Arbeitsagentur, LinkedIn, XING and career pages. Each needed its own visit.
- Stale postings. Aggregators keep showing roles the employer has already closed. Time spent on those is lost.
- Generic applications. One tailored application took me 1 to 2 hours or more. Under pressure, tailoring is the first thing that gets cut.
- Silence without a system. Many applications got no answer, and I could not tell which ones were worth chasing.
How might I hand the repetitive search and preparation work to AI agents, while keeping judgment, tone and every final decision with me?
My role and the rules I set
I designed the whole service: the flow, the decision rules, the scoring, the interface and the tone of every message. Claude, running in Cowork, is the worker that carries out those rules. Four constraints shaped every decision:
01
Human in the loop.
The system prepares, I decide and send. No application, email or message leaves without me.
02
Truth over polish.
Agents may only use facts from my CV and portfolio. No invented experience or metrics.
03
Respect the platforms.
Sites that block automated reading are never worked around. Roles found there are confirmed on the employer's own site instead.
04
Low maintenance.
It has to run while I am at the course or with my daughter, and only tell me what needs my attention.
Key design decisions
| Decision | Why | Trade-off I accepted |
|---|---|---|
| 1Verify every role on the employer's own site | A role I cannot confirm is a role I should not spend time on. | Some good roles on career sites that only load with JavaScript are listed as "worth a look" instead. |
| 2Two scores per role: Fit and Ready | Fit answers "should I apply?". Ready answers "is my application good enough to send?". A great match with a weak letter should not look ready. | Two numbers instead of one, so the card shows them as a ring and a small badge. |
| 3A weighted audit for every kit | Ready is a 7-part rubric: JD alignment 30, evidence 20, recruiter scan 15, ATS keywords 15, seniority 10, clarity 5, accuracy 5. 85 or more means send. | I adopted it from a friend's workflow, so it reflects one recruiter's view, not a universal standard. |
| 4Revise once, then stop | An agent can polish forever. One revision keeps runs short and predictable. | Some kits land just under 85 and need my own final edit. |
| 5A requirement map instead of a claim | 8 to 12 requirements from the posting, each rated strong, partial, not shown or can't claim, with evidence. | Longer to read, so it sits inside the expanded card, not on the list. |
| 6Follow-up rules after 21 days of silence | Each silent application is sorted into push for feedback, followed up or consider closing, with a reason and a channel. For the ones worth pushing, it drafts the follow-up email in Gmail for me to check and send. | The system suggests, I close. It never closes a role by itself. |
The tracker interface
The tracker is where the agents' work meets my decisions. I designed it as one working column: a pipeline strip that doubles as the filter, the latest digest, then role cards that expand into the full application kit. All screens use real data from my pipeline, with company names replaced and salaries removed.
01
Pipeline at a glance
The stage counts are the filter, so one tap answers "what needs me today?". The digest shows the top three and folds the rest away.
02
One role, one decision
The card leads with the honest view: why it fits, what is missing, and a requirement map rated against evidence from my CV.
03
The kit, ready to check
Keywords to mirror, plus the audit with its score and the improvements still open before I send.
04
Prepare on demand
I tick the roles I want and press one button. The Prepare Applications agent writes the tailored CV and its audit and tells me when it is done.
05
Follow-up, not guesswork
After 21 days of silence, each application gets a recommendation with a reason and a channel. I decide what to do with it.
The layout reflows on the phone to a two-column strip and full-width cards, so a quick check works anywhere.
Screens from 1 Oct 2026, before the cost redesign. Locations and identifying details are blurred, and one team count is corrected.
Iterations
5 Oct 2026
Redesign for cost.
Every agent run costs tokens, and tokens cost money. Usage grew faster than my budget allowed, so I paused all agents and redesigned: scheduled runs went from 8 to 4 a week, the digest stopped writing CVs, full CVs are only written on demand for roles I pick, and the search covers design roles only.
1 Oct 2026
Prepare Applications agent and the select-and-prepare button.
A full tailored CV and audit on demand, usually within about 20 minutes.
1 Oct 2026
Ready score and requirement map.
The quality of each kit became measurable.
1 Oct 2026
Job Inbox Check and follow-up rules.
The first review sorted 19 silent applications: 5 to push, 5 waiting, 9 to consider closing.
29 Sep 2026
UX Job Digest, the tracker, and an import of 66 roles from my manual search.
29 Sep 2026
Platform access test.
Before writing any search rules I tested which job sites can be read directly. LinkedIn, XING, Welcome to the Jungle and kununu block it, so roles from there are confirmed on the employer's site.
Results so far
Two weeks in, the clearest result is not volume but answers: when the system flags a silent application and drafts the follow-up, companies reply.
122
roles in one pipeline
3/4
follow-ups answered
1
interview invitation
As of 13 Oct 2026
- Roles in one pipeline
- 122 (66 imported from my manual search, 54 found by the agents, 2 added by hand)
- Applications sent since mid-August
- 78: 44 rejections, 33 without a final answer (sorted by the follow-up rules), 1 interview invitation
- Follow-ups on silent applications
- 4 recorded so far, 3 got a human reply
- Roles with a full application kit
- 30, fit scores 60 to 80 (average 69)
- Kits with a full audit and tailored CV
- 11, Ready scores 74 to 88, 3 reached 85 or more
- Scheduled agent runs per week
- 8 before the cost redesign, 4 after
What the follow-ups showed: I do not have many interviews yet, but I do get answers instead of silence. One follow-up the system flagged and drafted turned into an interview invitation from an AI company. They asked if I am at home in agentic AI, and this system was my answer. Another got a same-day reply: the rejection had been sent weeks earlier through their recruiting tool and never reached my inbox. A clear no is still an answer I can close and learn from.
Before: I found the role, checked it on the career page, researched the salary, worked out keywords with Claude over many rounds, then wrote and formatted the CV and letter. 1 to 2 hours or more. Now: the agents do that preparation, and an on-demand kit is usually ready in about 20 minutes of agent time while I do something else.
Process summary
How the work was made
Participants
Just me, with Claude as the worker, tested on 122 real job postings.
Result
A service I use every week. The follow-ups it flags and drafts get answers, and one of them became an interview invitation from an AI company that asked if I am at home in agentic AI.
What I would test next
Time and cost per application before and after, and whether agent-prepared applications get more replies than my manual ones.
Outcome
- One pipeline instead of a dozen job boards and inbox threads
- Every role the agents add is checked on the employer's own site first
- Two visible scores and an audit, so "is this good enough?" takes seconds to answer
- A follow-up rule for silence, with a reason and a channel for each application
- Half the scheduled agent runs after a cost redesign, with the same pipeline
What I took from it
Designing for an agent is designing a service: the brief, the rules and the hand-off points matter more than the screens. Every run starts from zero, so each brief has to stand alone. The human belongs where judgment lives: agents search, check and draft, while deciding, sending and closing stay with me. And cost is a design constraint too: every run has a price, so I designed for it like any other constraint, with fewer, better-timed runs and expensive work only on demand.
Next case study
E2E Return Optimization