Workflow design / Responsible AI / Decision support
Agentic Career Search
A Human-in-the-Loop Opportunity Intelligence and Application Workflow
I designed a human-in-the-loop agentic workflow that monitors new opportunities, applies evolving career criteria, compares job requirements with verified professional evidence, recommends next actions, and drafts tailored materials, while keeping every consequential decision under my control. It combines recurring monitoring, conditional logic, reusable criteria, tool-assisted research and drafting, and explicit review checkpoints. It does not apply for jobs on its own.
01 / The challenge
A serious job search is not a single prompt.
It is a changing decision environment: roles open and close, salary ranges shift, relocation and remote-work conditions matter, adjunct opportunities follow different rules from full-time roles, and new information can change whether an opportunity is still attractive. Repeating these judgments by hand across separate searches leads to inconsistent screening, duplicated work, stale assumptions, and applications that no longer reflect current priorities.
My design goal was to build a workflow that could remember the decision framework, watch for meaningful changes, gather evidence, adapt to revised criteria, and stop or continue based on explicit rules — while keeping final judgment with me.
02 / How the workflow evolved
More continuity. Clearer decision rules.
The project began as a human-coordinated workflow: I used ChatGPT for research, evaluation, and drafting, and Claude for refinement, moving work between them myself. It has since grown into a broader agentic system with four key additions:
Persistent decision criteria
Salary, role level, location, work arrangement, strategic value, adjunct economics, and economic-equivalence exceptions are stored as reusable screening rules instead of being re-entered for every search.
Recurring opportunity monitoring
A scheduled job watch checks target roles and searches for newly posted openings, surfacing only meaningful changes or strong matches.
Conditional branching
Each opportunity is routed to a decision path (full-time, economic-equivalence exception, adjunct/part-time, or strategic academic) and given a recommended next action: Apply, Pass, Clarify, or Monitor.
Feedback and correction loops
When I revise a criterion or reject an assumption, I update the working criteria so that future recommendations apply the corrected rule.
03 / Workflow architecture
Eight stages, with bounded autonomy.
01Define
Agentic functionCareer criteria, target role families, compensation strategy, location flexibility, adjunct rules
Human checkpointHuman sets priorities and boundaries
02Monitor
Agentic functionScheduled checks for target-role changes and new matching opportunities
Human checkpointAI searches; human sees only meaningful changes
03Research
Agentic functionCurrent posting, employer source, compensation, location, deadline, requirements
Human checkpointAI gathers evidence; human can request deeper verification
04Evaluate
Agentic functionJob requirements + résumé/CV evidence + evolving criteria
Human checkpointAI distinguishes strong fit, transferability, gaps, unknowns
05Decide
Agentic functionApply / Pass / Clarify / Monitor, within the matching decision path
Human checkpointHuman accepts, questions, or overrides the recommendation
06Produce
Agentic functionTailored résumé/CV, cover letter, salary strategy, outreach, interview prep
Human checkpointAI drafts from verified evidence; human reviews and approves
07Refine
Agentic functionOptional second-model editing and targeted revisions
Human checkpointHuman coordinates handoffs and checks meaning/accuracy
08Learn
Agentic functionOutcomes, corrections, new constraints, changing priorities
Human checkpointHuman updates criteria and prompts for future cycles
04 / A living decision model
Different opportunities. Different paths.
The criteria are deliberately open to revision. The workflow treats different opportunity types differently and lets the rules evolve as the search produces new information.
Primary full-time path
Senior instructional design, learning experience design, academic technology, digital/online learning leadership, faculty development, learning technology leadership, accessibility/quality, performance consulting, organizational learning, applied analytics, and AI-enabled learning. Compensation is evaluated against the current full-time strategy.
Economic-equivalence exceptions
A lower base can still merit consideration when several advantages coexist, such as excellent benefits, manageable workload, remote/hybrid flexibility, schedule control, outside-work potential, career-growth value, dual-career opportunities, or relocation support.
Adjunct / part-time path
Evaluated separately from the full-time search, with compensation, remote delivery, preparation burden, subject alignment, and strategic academic value weighed differently.
Strategic academic path
A lower-paying adjunct or faculty-pool opportunity may still be worthwhile when it adds graduate-level teaching, strengthens my positioning in AI and instructional technology, brings a strong institutional affiliation, or adds portfolio value.
05 / Human judgment in practice
Every consequential action is my decision.
Stopping is a valid output
If a posting fails key constraints, the workflow ends at screening, and no time goes into materials for a role I will not pursue.
Unknowns are not treated as facts
Missing salary, remote-work rules, relocation support, or appointment conditions are flagged for clarification, not guessed.
Criteria can be corrected
When a recommendation rests on an outdated assumption, I revise the rule, and later evaluations use the updated distinction.
Gaps are reframed honestly
For a learning-delivery role in manufacturing, screening flagged a language-proficiency requirement I did not meet. I chose to apply anyway, presenting my multilingual background, linguistics and language-education expertise, and international experience as transferable strengths, without claiming the missing qualification.
External actions remain controlled
Applications, emails, salary commitments, and other consequential actions require my explicit decision.
06 / Tools and role separation
Defined contributions. Explicit boundaries.
ChatGPT
Primary contributionResearch orchestration, opportunity evaluation, recurring monitoring, evidence comparison, drafting, decision support
BoundaryDoes not make final career decisions or submit applications without approval
Claude
Primary contributionOptional refinement and polishing of selected drafts; an alternative editorial perspective
BoundaryA human-directed second pass; does not communicate with other tools on its own
Career criteria / source documents
Primary contributionProvide stable evidence, constraints, and prior decisions
BoundaryMust be reviewed when circumstances change
Human reviewer — me
Primary contributionDefines goals, resolves ambiguity, corrects assumptions, approves outputs, decides when to stop or proceed
BoundaryRetains accountability for all consequential decisions
07 / Why I describe this as agentic
Agentic, not unchecked.
The workflow shows several characteristics of agentic systems: persistent goals, task decomposition, recurring monitoring, tool use, evidence retrieval, conditional next steps, reusable context, feedback loops, and multi-step execution. The key distinction is that its autonomy is bounded: the system performs recurring research and produces recommendations, but consequential decisions remain human-approved.
Precise description: a human-in-the-loop agentic workflow with scheduled monitoring and human-coordinated multi-assistant refinement — not autonomous multi-agent hiring or application submission.
08 / Outputs produced
Evidence, decisions, and tailored materials.
- Current-role status checks and new-opportunity discovery
- Fit assessments with evidence-based Apply / Pass / Clarify / Monitor recommendations
- Salary-expectation and economic-equivalence analysis
- Targeted résumés, academic CVs, cover letters, outreach messages, and interview preparation
- A living job-search criteria playbook with defined decision paths
- A documented trail of corrections, exceptions, and revised rules that improve later decisions
09 / Evaluation plan
Measuring the workflow.
The next stage is to measure the workflow rather than rely on qualitative impressions alone. I have begun tracking:
- Time from opportunity discovery to Apply / Pass decision
- Number of irrelevant opportunities filtered before document preparation
- Interview rate by opportunity category
- False positives and missed opportunities
- Frequency of human corrections to AI assumptions
- Reuse of verified evidence across tailored applications
- Whether recurring monitoring surfaces opportunities I would otherwise have missed
10 / Transfer to learning and development
A prototype for responsible agentic design.
The same architecture transfers to instructional design work. A course-quality or performance-consulting workflow could monitor incoming needs, retrieve relevant standards and evidence, classify the problem, recommend an appropriate intervention, draft artifacts, route uncertain cases for human review, and learn from reviewer corrections. The job-search case is therefore both a personal productivity workflow and a prototype for responsible agentic design in learning and development.
Key takeaway
Agentic value does not require removing the human.
The most useful improvement in this project came from giving the system more continuity, clearer decision rules, recurring work, and explicit feedback loops — while preserving human judgment wherever the consequences matter.