Platform · agents · release 2026.4
Thirteen agents, one intelligence layer, inside your VPC.
Every Nodes agent reads and writes to the same calibrated scoring substrate — the one trained on your own production outcomes. Agents specialize; the model underneath doesn't split. Source, interview, ramp, retain, and develop, all traced to the same evidence.
The agent roster
Four stages of the talent lifecycle. One scoring substrate.
Agents are not separate models stitched together. They are specialized surfaces on top of the same calibrated weights — carrier-hire-2025.11 — so a signal seen in sourcing compounds into the retention stream ninety days later.
Stage · 01
Pre-hire
finding & screening signal
against production pattern
against production pattern
agent · 01live
Persona Sourcing
Queries learned patterns against LinkedIn, GitHub, industry networks. Ships provisional Fit Scores.
agent · 02live
Fit Screening
0–100 ATS-native score across 28 behavioral, skill, and cultural dimensions.
agent · 03live
Pipeline Triage
Dedupes, re-ranks, and routes requisitions by req-level calibration drift.
Stage · 02
Interview
structured signal collection
with signed evidence
with signed evidence
agent · 04live
Structured Interviewer
Async or live mid-funnel conversations scored on six canonical signals. 12-min median.
agent · 05live
Reference Auditor
Verifies claimed outcomes by triangulating structured reference calls and public track record.
Stage · 03
Post-hire
ramp, retain, forecast
against real outcomes
against real outcomes
agent · 06live
Ramp Acceleration
Role-calibrated onboarding signal. Scored hires reach first milestone 47 days faster.
agent · 07live
Retention Risk
Early-warning signal stream over behavioral and production trace data. Weekly refresh.
agent · 08live
Attrition Modeling
Cohort-level forecasting, calibrated monthly against HRIS outcomes.
agent · 09live
Manager Intelligence
Team-level production trace — surfaces what actually correlates with manager-driven lift.
agent · 10live
Team Composition
Portfolio view of trait coverage across a squad; flags concentration risk on key dimensions.
Stage · 04
Develop
mobility, succession,
personalized coaching
personalized coaching
agent · 11live
Internal Mobility
Cross-role pattern matching on the same dimensional fingerprint used at hire.
agent · 12live
Succession Planning
Key-role readiness scoring, persona-calibrated against validated top-performer traces.
agent · 13beta
Mentor Co-pilot
Personalized coaching for managers and ICs grounded in the same decision trace.
avg lift2.47×
time-to-hire127d → 38d
inference p5074 ms
data egresszero
Anatomy of an agent
Inputs on the left. Model in the middle. Actions on the right.
Every agent is a thin orchestration layer over the same calibrated substrate. Swap the prompt, not the model. Swap the output surface, not the weights. That's why an interview score and a retention score agree — they read the same evidence.
Inputs · evidence
What the agent reads
ATS candidate recordreq · resume · stage▸
- source
- Workday · Greenhouse · iCIMS
- fields
- req_id, stage, resume_parsed, recruiter_notes
- cadence
- webhook · sub-second
Structured interview transcriptsigned · timestamped▸
- source
- agent · 04 · structured interviewer
- payload
- utterance stream, 6 signal flags, sig-hash
- retention
- customer-vpc · indexed by req_id
HRIS production stream18-mo rolling▸
- source
- UKG · ADP · Workday HCM
- signal
- RPA, premium-per-month, retention event
- window
- rolling 18-mo · refresh weekly
CRM pipeline eventsper-agent cadence▸
- source
- Salesforce · Dynamics · HubSpot
- signal
- activity cadence, pipeline velocity, conversion
- grain
- per-agent · per-week
Persona query resultlearned pattern▸
- source
- agent · 01 · persona sourcing
- payload
- candidate + provisional fit + source attribution
- pattern
- learned · not keyword
Reference + audit trailsig-hash chain▸
- source
- agent · 05 · reference auditor
- payload
- structured ref calls · public filings triangulation
- chain
- hash-linked to screening + interview traces
Substrate · one model
The calibrated weights
model.id · carrier-hire-2025.11
carrier-hire-2025.11
fine-tuned · customer-vpc · quarterly retrain
dimensions
28
corpus
10,765
auc · rpa
0.618
p50 · infer
74ms
orchestration · agent wrapperthin
prompt · signal mapper-agent
calibration · role · locationreq-level
policy · boundariescustomer-owned
Outputs · actions
What the agent writes back
Fit Score · 0–100ats native field▸
- writes to
- ATS candidate.fit_score (native field)
- calibration
- per-role · per-location · req-level
- shape
- integer 0–100 · 28-dim breakdown · rationale
Rationale · plain englishreviewer-facing▸
- writes to
- ATS candidate.rationale · reviewer view
- content
- which dims drove the score · top 3 · bottom 3
- length
- 3–5 sentences · never hedged
Decision tracesigned · reproducible▸
- writes to
- audit stream · SIEM export
- signature
- sig-hash · per-req · chain-linked
- replay
- reproducible on quarterly retrain
Signal streamweekly · webhook▸
- writes to
- manager inbox · HRIS · BI warehouse
- cadence
- weekly · delta since last emission
- payload
- retention risk, ramp drift, coaching flag
Cohort forecastmonthly rollup▸
- writes to
- Snowflake · Databricks · BigQuery
- grain
- cohort × role × location · p10/p50/p90
- calibration
- reweighted against outcomes monthly
Recommended actionmanager inbox▸
- writes to
- manager email · Slack · Teams
- content
- one action per agent · grounded in trace
- policy
- customer-owned thresholds · no auto-fire
Agents in concert
One requisition, six agents, one trace.
Requisition rv-2412-08831 — Commercial Lines Producer, Austin. This is the actual sequence agents execute, end to end. Every output below is signed and writes back to the same trace chain.
trace · rv-2412-08831 · producer.commercial.austin
day 0req opened
day 2sourcing
day 6screening
day 14interview
day 34offer accepted
day 62post-hire
01
agent · 01 · persona sourcing
Materialized 218 passive candidates from LinkedIn, GitHub, and regional industry networks against the
top_producer_commercial_lines_austin persona.218 candidates4 sourceslearned pattern
provisional fit74 +
02
agent · 03 · pipeline triage
Deduped against the active pipeline, re-ranked by req-level calibration, routed 42 to screening and 12 to direct-sourcing outreach.
dedupe 38routed 54ats writeback
queued54
03
agent · 02 · fit screening
Scored 54 candidates across 28 dimensions. Top of funnel: M. Rojas, score 87 — p-74 vs. the austin-commercial cohort. Rationale and trace written back to Workday.
28 dimstrace signedworkday native
top score87
04
agent · 04 · structured interviewer
Conducted 12-minute async interview. 4 of 6 canonical signals present (resilience, process reset, self-audit partial, prospecting drive). Transcript signed, hash chained to screening trace.
12:04 recorded4/6 signalssig hash · 0a91f…d72e
advance87
05
agent · 05 · reference auditor
Triangulated claimed production outcomes at two prior carriers against public filings and three structured reference calls. No reconciliation exceptions.
3 refs · structured0 exceptions
verifiedclean
06
agent · 06 · ramp acceleration
Handed off to ramp agent at day 34. Role-calibrated onboarding plan scheduled against the austin-commercial 62-day SNA median. Retention signal stream activated on week 2.
sna target · d62retention stream · onmanager briefed
milestoned62
one model · one trace · six agents
req closed · d34 · trace · dc_8871_2026-04-14
Governance
Every agent inherits the same boundary.
No agent is a SaaS call. No agent loads a third-party model. The substrate runs inside your cloud account, retrained on your data, audited against your controls.
Deployment
Customer VPC, single-tenant
AWS · Azure · GCP. Terraform module ships at contract.
Model weights
Owned by the customer
Fine-tune runs inside the cluster. No external checkpoints.
Data egress
Zero, by architecture
No candidate PII, transcripts, or embeddings leave the boundary.
Audit
Signed decision trace
SIEM stream, hash-chained per req, reproducible on retrain.
See the thirteen agents run against your own four years of data.
A 30-minute working session. We model the cohort, score a hold-out, and show the decision trace for three hires you already made.