Skip to content

The architecture

32 specialized agents.
One career assistant.

Most AI job tools ship one chatbot in a trench coat. NovaHunt is a tier hierarchy of small, focused agents — each one auditable, each one with a hard spend cap, each one doing exactly one job.

The constellation

All 32 agents, at a glance.

VANGUARD at the center. Eight directors radiate out. Twenty-one specialists cluster near their director. Three safety agents (dotted lines) govern every call.

Sourcing

HARVESTER · POACHER · DRAGNET · SIEVE · CRIER

Scoring

PARAGON · DOWSER · APPRAISER · JUDGE · CALIBRATOR

Learning

SCOUT

Documents

SCRIVENER · REDACTOR · OVERTURE · MUSE · PERSONA · STEWARD

Applying

EMISSARY · PORTAL · DISPATCHER · WITNESS

Follow-up

ENVOY · CADENCE · LISTENER

Strategy

STRATAGEM · CARTOGRAPHER · REFLECTOR · AUDITOR · EXPLORER

Safety layer

WARDEN · TURNSTILE · ABACUS

Orchestrator

VANGUARD

Sourcing · 5 agents
Scoring · 5 agents
Learning · 1 agent
Documents · 6 agents
Applying · 4 agents
Follow-up · 3 agents
Strategy · 5 agents
Safety layer · 3 agents

VANGUARD at the center orchestrates. Eight domain directors radiate out in the inner ring. Twenty-one specialists cluster near their director. Three safety agents (dotted lines) govern every call. Read the architecture →

The design principle

One monolithic LLM that does everything is impossible to audit, impossible to bound, and impossible to improve. NovaHunt splits the job into 32 focused agents organized in three tiers. Each one is independently versioned, independently rate-limited, and independently swappable. If APPRAISER's scoring drifts, we replace just APPRAISER — not the whole assistant.

Underneath them sits a deterministic safety layer (WARDEN, TURNSTILE, METER, CANARY) that is written in pure Python and contains zero LLM imports. Agents propose, the safety layer disposes. That's what makes the spend caps real.

Tier 1

The conductor

One agent that orchestrates every other agent and assembles your daily brief.

VANGUARD

Conductor
Sees

Every agent's queued briefing items from the last 24h

Produces

Your daily brief: hot matches, follow-ups owed, replies that need you

Cost ceiling

$0.05/run

Tier 2 — 8 directors

Domain directors

Each one owns a stage of the search pipeline and delegates to its specialists.

HARVESTER

Sourcing Director
Sees

Your target roles, locations, and enabled job boards

Produces

Newly fetched + deduped postings across every connected source

Cost ceiling

$0 (no LLM — pure HTTP)

PARAGON

Matching Director
Sees

Newly sourced postings and your profile

Produces

Scored, ranked, and rationale-attached matches

Cost ceiling

$0.50/batch (max 12 jobs)

SCRIVENER

Documents Director
Sees

A specific job description and your master profile + voice

Produces

Tailored resume + cover letter as new document versions

Cost ceiling

$0.10/job

EMISSARY

Application Director
Sees

A drafted application and your channel preference

Produces

Queued ATS submission or composed recruiter email

Cost ceiling

$0.04/application

ENVOY

Follow-up Director
Sees

Every application's stage and time-since-submit

Produces

Overdue-follow-up flags for VANGUARD to escalate

Cost ceiling

$0 (no LLM)

STEWARD

Assets Director
Sees

Your resume + cover letter library and last-edit timestamps

Produces

Staleness flags, version recommendations, missing-template warnings

Cost ceiling

$0 (no LLM)

STRATAGEM

Strategy Director
Sees

Your full search history week-over-week

Produces

Pattern-spotting briefings ("your offers come from sub-50 person companies")

Cost ceiling

$0.05/week

CARTOGRAPHER

Coverage Director
Sees

Which sources produce your highest-scoring matches

Produces

Source-mix recommendations and gap-coverage suggestions

Cost ceiling

$0 (no LLM)

Tier 3 — 21 specialists

Focused specialists

Each does exactly one thing — score a job, write a cover letter, classify a reply — and does it cheaply.

POACHER

ATS specialist
Sees

Public ATS endpoints (Greenhouse, Lever, Ashby, Workable)

Produces

Raw postings normalized to one schema

DRAGNET

Aggregator specialist
Sees

Adzuna + JSearch (LinkedIn/Indeed/Glassdoor proxy)

Produces

Raw postings normalized to one schema

SIEVE

Dedup specialist
Sees

Raw postings from every source this run

Produces

Deduped jobs keyed by (source, source_id) + URL hash

CRIER

Notifier
Sees

Hot matches and digest payloads

Produces

Delivered email via Resend MCP

DOWSER

Prefilter specialist
Sees

Candidate jobs vs your profile vector

Produces

Top-K shortlist when the candidate pool > 50

APPRAISER

Per-job scorer
Sees

One job + your profile

Produces

Normalized score [0,1] + 3-5 weighted factors

Cost ceiling

$0.02/job

JUDGE

Rationale specialist
Sees

An APPRAISER score above 0.60

Produces

Plain-English rationale + gaps + top-applicant tips

Cost ceiling

$0.04/job

SCOUT

Local learner
Sees

Your saved + dismissed labels (sklearn TF-IDF + LogReg)

Produces

Personalized fit score that re-ranks the list — $0 marginal cost

CALIBRATOR

Quality auditor
Sees

Scored matches vs your real saves/dismisses

Produces

Score-calibration drift signals

REDACTOR

Resume tailorer
Sees

Your master resume + this specific JD

Produces

Tailored resume emphasizing JD keywords

Cost ceiling

$0.04/job

OVERTURE

Cover-letter writer
Sees

JD + your achievements + voice notes

Produces

Cover letter in your tone

Cost ceiling

$0.04/job

MUSE

Voice analyst
Sees

Your uploaded resumes + cover letters

Produces

Voice fingerprint so future drafts sound like you

PERSONA

Top-applicant brief
Sees

This JD + your profile

Produces

Archetype + ✓/○ checklist + differentiator

Cost ceiling

$0.03/job (cached)

PORTAL

ATS submission staging
Sees

Drafted application + ATS metadata

Produces

Hands you a deep-link to submit (full automation on roadmap)

DISPATCHER

Email submission specialist
Sees

Drafted application + recruiter address

Produces

Sent email via Resend with your reply-to

WITNESS

Confirmation logger
Sees

Confirmation reference (ATS ref / message-id)

Produces

Verified application stage flip to 'applied'

CADENCE

Follow-up drafter
Sees

Submitted application + days elapsed

Produces

Tone-appropriate follow-up email draft

Cost ceiling

$0.02/draft

LISTENER

Reply classifier
Sees

A recruiter's inbound reply

Produces

Classification (interview / rejection / etc.) + extracted action

Cost ceiling

$0.02/reply

REFLECTOR

Outcome analyst
Sees

Applications and their final stages

Produces

Conversion-rate analysis per role / per source

Cost ceiling

$0.05/week

AUDITOR

Process auditor
Sees

Every agent_run for the week

Produces

Spend + latency + failure breakdown

EXPLORER

New-source scout
Sees

Source performance + jobs you missed elsewhere

Produces

Recommendations for new boards to enable

ABACUS

Spend anomaly detector
Sees

Hourly + daily spend rollups across every user

Produces

Anomaly flags that halt spending until you approve

The safety layer

Deterministic guardrails

Zero LLM imports. The safety layer is what makes the spend caps and consent tiers actually enforceable — an agent cannot reason its way past a deterministic Python validator.

WARDEN

Deterministic safety validator

Zero LLM imports. Pure Python validators that gate every agent action against the Constitution. If WARDEN says no, the action never happens — the agent can't talk it into yes.

TURNSTILE

Pre-spend authorization gate

Every Claude call asks TURNSTILE first. It checks your daily and monthly caps, your plan tier, your active spend holds, and your consent tier for that action class. Denial returns a 402 'approval needed' — the agent halts cleanly.

METER

Token + cost ledger

Records every Anthropic call (model, input tokens, output tokens, USD) into the agent_runs table. Drives your spend dashboard and ABACUS's anomaly detection.

CANARY

Event-rule monitor

Watches the audit stream for suspicious patterns (e.g. an agent that suddenly spends 10× its usual rate) and emits halt-and-approve events to the user.

WATCHTOWER

Per-user resource governor

Tracks rate limits per action class (intake parses/hour, refreshes/hour, auto-applies/day) and routes denial through TURNSTILE so the user sees the same '402 approval needed' UX everywhere.

Consent tiers

GREEN / YELLOW / RED — you decide what's automatic

Every action class has a consent tier. GREEN actions run silently (e.g. extract profile from a resume you just uploaded). YELLOW actions queue for your one-tap approval (e.g. submit an application). RED actions require step-up authentication (e.g. rotate your API credentials). You can re-tier any action at any time in Settings → Consent; RED actions can never be auto-promoted to GREEN.

See it run on your own resume

Sign in, upload your resume, and the agents go to work. First refresh surfaces real matches; first save trains your personalized re-ranker. No card required to try.