◆ Assistive Worker
Scores every application against the role's competency rubric (semantic match, not keyword filter) and assembles a ranked shortlist with per-call evidence. Ranks and evidences only, never rejects autonomously; a judge pass audits fairness on a sampled basis and the hiring-decision agent acts on the shortlist.
Memory
Working The role rubric and the candidate currently being scored.
Episodic Prior applications from the same candidate across requisitions.
Semantic Competency frameworks, role families, must-have vs. nice-to-have criteria.
Procedural Calibration learned from the hiring-decision agent's accept/reject feedback.
Store Vector store over the ATS + structured rubric library
Orchestration
pipeline MCP
Harness · Managed Agents … session per requisition; context editing drops stale résumé text once a candidate is scored.
Tools
{ } Applicant tracking system (ATS) API ›_ Résumé parsing Code exec ⌕ Rubric retrieval Retrieval ⇄ Hiring-decision agent handoff A2A
Evals & guardrails
- Mandatory bias/fairness checks: adverse-impact analysis across protected groups every cycle.
- The downstream hiring-decision agent acts … this agent only ranks and evidences, never rejects autonomously.
- Agent-as-judge audit of ranking rationale against the rubric on a sampled basis.
- Immutable log of every score for EEOC defensibility.
Offline reflection
Consolidates the hiring-decision agent's overrides offline to recalibrate the rubric weighting … Reflexion-style, with a fairness regression-check before any change ships.
Frontier edge
- ▲Eval-gated continual learning (SEAL-style): every hiring-decision-agent accept/reject nudges rubric calibration without a full retrain, but only after the adverse-impact regression passes.
- ▲Counterfactual fairness probing: re-scores each candidate with protected attributes perturbed to catch proxy bias before a shortlist ships.
- ▲Confidential compute: résumé PII stays inside a secure enclave; the decision agent sees scores and evidence, never raw data in clear.
A sample run
Trigger 320 applications land for a quant developer requisition.
- 1Score each résumé against the competency rubric, not keywords.
- 2Cluster near-duplicate skill profiles; surface non-obvious adjacent experience.
- 3Run the adverse-impact check across the candidate pool.
- 4Assemble a ranked top-25 with per-candidate evidence.
Output A ranked shortlist with rationale handed to the hiring-decision agent, which decides who advances under the fairness judge. Nothing is auto-rejected by this agent.
In numbers
6,800
Résumés screened / day
under 4 min
Median time to shortlist
tracked per requisition
Shortlist→interview acceptance
Handoffs
Hands to → Interview Scheduling Agent