Beyond the AI Hiring Doom Loop - Part II
From single agent Q&A to recruiter candidate conversations
In my previous artilce on the AI hiring doom loop, I described a pattern we are starting to see more often. Candidates use AI to produce perfectly optimized resumes. Companies use AI to screen them at scale. The result is that the systems can end up reinforcing each other. Not improving signal, but amplifying noise.
One idea I proposed was to move away from “optimized resumes” and toward AI agents that represent candidates in a grounded way, based on real experience and curated evidence. This follow up is about the next step. A structured conversation between a recruiter agent and a candidate agent. Not a static form. Not a keyword scan. A short screening conversation that leaves a trace.
This is an early prototype, but it already shows a viable path forward.
The core idea
Instead of reading yet another perfect resume, a recruiter gets three things:
A structured professional summary tailored to the role
A short transcript of the screening conversation
A transparent scorecard with the rationale behind it
And the candidate gets something rare in modern hiring: A record of the interaction and feedback that can be reviewed, instead of disappearing into the ATS black hole.
The architecture, in plain terms
I started from an existing candidate agent, my Professional Digital Twin, and paired it with a recruiter agent designed to run a consistent first screen. The system is built using the OpenAI Agents SDK in Python, with built in session management, tracing, and conversation history. There are three components:
1) Candidate agent Grounded in a curated knowledge base containing real experience, achievements, and professional background.
2) Recruiter agent Starts the conversation, asks targeted follow up questions, and produces an assessment using a scoring rubric.
3) Orchestrator A simple Python workflow that routes the turns between agents, tracks state, and persists sessions so each run can be inspected and audited.
The conversation flow
The prototype runs in four phases:
1) Initiation - The recruiter agent requests a brief, role tailored professional summary.
2) Initial response - The candidate agent produces the summary using only its knowledge base. No generic internet knowledge, no invented details.
3) Follow up dialogue - The recruiter agent asks two to three targeted questions, focused on typical first screen signals:
Experience with key technologies or methods
Leadership style and team management approach
Domain specific achievements and challenges
Alignment with constraints like location, seniority, and work mode
4) Final assessment - At the end, the recruiter agent produces a structured JSON evaluation, including:
Skills score, 0 to 100
Domain fit score, 0 to 100
Constraints alignment score, 0 to 100
Overall recommendation, proceed or do not proceed
The key point is that the output is not just a score. It is a conversation plus rationale that can be reviewed, audited, and improved.
Why traceability matters
During execution, every turn is stored with timestamps. Candidate and recruiter have separate sessions linked by the orchestrator. This makes it easy to inspect a specific run, understand why the system reached a decision, and spot patterns over time. If we want to trust agents in hiring, this kind of visibility is not optional.
Challenges so far
This is a PoC built to explore the strengths and limits of the approach. A few takeaways stood out.
1) Prompt dependency is real
Both agents are heavily shaped by their system prompts. Vague instructions lead to meta questions and clarifications. Overly rigid output requirements can cause the recruiter to rush to scoring. If you are not explicit, agents can drift out of role or answer too generically.
Scaling across roles becomes a prompt and UX problem as much as an engineering one.
2) The orchestrator is intentionally simple
That is good for predictability and debugging, but it limits realism.
Fixed number of question
Linear flow
No branching
No negotiation on constraints
Limited adaptation to different profiles
3) Grounding depends on knowledge quality
The candidate agent is only as strong as its knowledge base and retrieval. If the knowledge is incomplete or retrieval surfaces the wrong material, the answers degrade. The upside is that it fails safely. It says “I do not have enough information” rather than hallucinating. But it reinforces a key lesson: knowledge curation is part of the product.
Where this goes next
A production ready version would likely need:
MCP based knowledge management A managed, updateable layer for candidate profiles, role requirements, and recruiter guidelines, with versioning and controlled sharing.
Smarter orchestration Dynamic stopping rules, branching flows, deeper probing, and context aware follow ups.
Output validation Schema validation and automatic retries to make structured assessments reliable enough to integrate with ATS, CRM, and analytics.
Multi persona support Different recruiter styles, role specific agents, and even panel style evaluations where multiple recruiter agents assess the same candidate from different angles.
The bigger point
This is an early step toward shifting from resume based filtering to conversation based evaluation.
Done right, recruiter candidate agent conversations can:
Preserve signal quality by grounding responses in real evidence
Scale without reducing everything to keyword matching
Provide transparency through transcripts, rationale, and traceable scoring
Keep humans in control, with better inputs rather than opaque filters
The AI hiring doom loop is not inevitable. If we design for authenticity, transparency, and structured evaluation, we can build something better for both candidates and hiring teams.