Nora Architecture

NAGENT AI · AI SALES TEAM · TECHNICAL SPECIFICATION NORA — Agent architecture Orchestrator + six sub-agents · skills & tools per agent · from the 25 Aug enhancement flow Document owner: Revenue & Sales · Version 1.0 · 25 Aug 2026 · Source: NORA enhancement whiteboard (25 Aug) + shipped product behaviour. Skills are typed prompt or logic — logic skills are deterministic and testable; prompt skills are grounded in the Knowledge Base. 1 · Architecture at a glance NORA is one product with one workflow, executed by an orchestrator and six specialised sub-agents. The orchestrator owns state and sequencing; sub-agents own one job each. Everything outbound — every e-mail, LinkedIn message and WhatsApp message, and every CRM write — passes through the approval gate, which is architecture, not a setting. Full-resolution version of this diagram is delivered alongside as NORA-agent-architecture.png. 2 · How a campaign runs — the workflow Order of operations: PLAN builds understanding and the campaign checklist → DISCOVER builds the company longlist and cuts it to a shortlist against ICP 1·2·3 → ENRICH researches each shortlisted company and its 5-person buying committee → ENGAGE drafts and (after approval) sends across e-mail, LinkedIn and WhatsApp, qualifies replies and books meetings → ANALYST reports the funnel weekly → ONBOARD activates customers who close, and feeds what it learns back into PLAN's Knowledge Base. Funnel defaults (set in PLAN's checklist, user-overridable): 100 companies 80 qualified 15 shortlist ×5 people ≈ 75 contacts longlist vs ICP 1·2·3 triggers attached, deduped vs CRM deep research, dossier each buying committee mapped & sequenced

-- 1 of 8 --

Two base tools ship in every sub-agent: deep research and web search — plus read access to the Knowledge Base. Everything else below is agent-specific.

-- 2 of 8 --

3 · NORA — the Orchestrator (1st Agent) NORA herself is the workflow owner, not a worker. She holds the campaign state machine, dispatches sub-agents in order, enforces the approval gate on everything outbound, and narrates status to the user. Sub-agents never talk to the outside world directly — they return work to NORA. Input A campaign brief (one line or a saved ICP profile) + the customer's company & product Output A running campaign: booked meetings, weekly analysis report, full audit trail Definition of done Meetings on a human calendar with the dossier attached; no unapproved send ever Skills Skill Type What it does Workflow orchestration Logic Runs the PLAN → DISCOVER → ENRICH → ENGAGE → ANALYST sequence; parallelises research; retries failed steps Campaign state machine Logic Every account has exactly one stage; transitions are logged and reversible Approval routing Logic Collects every outbound draft and proposed CRM write into the approval queue; blocks until human decision Escalation (36 h rule) Logic Any qualified handoff nobody acknowledges within 36 hours is escalated, loudly Status narration Prompt Explains what she is doing and why in the workspace channel and Ask-NORA chat Tools Sub-agent dispatcher Spawns and sequences the six sub-agents with scoped context Workspace state store (Smriti) Single shared memory across all sub-agents — accounts, stages, decisions Approval queue The gate: Approve / Edit / Skip per item, bulk approval for first-touch batches CRM read/write (gated) Salesforce · HubSpot · Zoho — writes queue as proposals until approved Notifier Slack / Teams / in-app for approvals, escalations and the Friday report Guardrail: Autonomy is earned, not deployed — the level (L1→L5) widens what sub-agents may do alone, never whether the gate exists. 4 · PLAN — understanding & development (Sub-Agent 1) The starting point from the whiteboard. PLAN turns the input company (its ICP and product) into two artefacts every other agent runs on: a stocked, source-tagged Knowledge Base and the campaign checklist. Input The customer's company: website, product, ICP hypotheses, existing collateral Output Campaign checklist (targets, ICP 1·2·3 definitions, channel plan, funnel defaults) + KB shelves: pricing · sales deck · e-mails/comms · competitive intelligence · sales material Definition of done Every checklist item has an owner (sub-agent) and every KB claim has a source Skills

-- 3 of 8 --

Skill Type What it does ICP & product understanding Prompt Reads the company's site, deck and docs; drafts ICP 1·2·3 and the value narrative for human sign-off KB curation Prompt Files pricing, sales deck, e-mail templates, competitive intel and sales material onto shelves with source tags and refresh rules Checklist generation Logic Compiles the campaign checklist — targets, funnel defaults (100→80→15, ×5 people), channels, cadence Gap detection Logic An empty shelf produces a question to the user, never an invention Tools Doc & site parser PDF, PPTX, DOCX and website crawl into structured KB entries KB writer (source-tagged) Every entry carries its source and a refresh rule; stale entries get flagged Checklist engine Template + overrides; the checklist is the contract for the whole campaign Guardrail: No source on file, no claim in any downstream message — PLAN's KB is the only place ENGAGE may draft from. 5 · DISCOVER — longlist → shortlist (Sub-Agent 2) Builds the target-company long list against ICP 1·2·3 and runs the first two funnel cuts: 100 longlist → 80 qualified → 15 shortlist. Output is a ranked shortlist where every company carries a written reason to exist on it. Input PLAN's checklist (ICP definitions, exclusions, geography, funnel sizes) Output 15 shortlisted companies, each with ICP tag, trigger, and ranking rationale Definition of done Zero duplicates vs CRM; every shortlisted company has a named trigger Skills Skill Type What it does ICP-match scoring Logic Need × Fit rubric per ICP; scores are numeric, repeatable and auditable Trigger detection Prompt Funding, hiring, leadership change, expansion — read from news, job posts and filings Dedupe & suppression Logic Removes existing customers, open opportunities, do-not-contact domains from the CRM Shortlist ranking Prompt Orders the 15 with a one-line why-now per company for human review Tools Company search / database API e.g. Apollo company search & enrichment for the longlist News & job-postings feeds Trigger evidence with links CRM read (suppression) Existing accounts and do-not-contact lists Guardrail: The 100/80/15 numbers are defaults from the flow — PLAN's checklist can override them per campaign. 6 · ENRICH — company + people research (Sub-Agent 3)

-- 4 of 8 --

The two red bubbles on the whiteboard: research on the target company, and people research. For each of the 15 shortlisted companies ENRICH maps the 5-person buying committee and writes the cited dossier that makes every later draft defensible — at roughly $0.23 of research per account. Input The 15-company shortlist with ICP tags and triggers Output ≈75 validated contacts + one dossier per account: fit reasoning, 4 pitch hooks, buying signals with links, committee map Definition of done Every dossier claim has a source link; every contact has a validated channel Skills Skill Type What it does Buying-committee mapping Logic + Prompt Which 5 roles matter for this ICP (logic), and who holds them here (research) Cited dossier writing Prompt Per-claim sources; refuses to state what it cannot source Pitch-hook extraction Prompt 4 hooks per account, each tied to a sourced signal Contact validation Logic E-mail verification, LinkedIn profile match, channel preference Tools People search & enrichment API e.g. Apollo people search / match for committee contacts LinkedIn lookup Role confirmation and recent activity Dossier store Writes into the account's CRM record; visible in the console's account view Guardrail: Research spend is metered and shown per account — the receipt trail is a feature, not overhead. 7 · ENGAGE — outreach, approval, booking (Sub-Agent 4) The left column of the whiteboard: E-mail, LinkedIn message, WhatsApp — and the circled word under them, Approval. ENGAGE drafts everything, sends nothing on its own, qualifies replies through the state machine and lands the booked meeting with the dossier attached. Input ≈75 contacts + dossiers + KB (the only permitted source of claims) Output Approved multi-touch sequences live across 3 channels; qualified replies → booked meetings, warm-handed to a human Definition of done Meeting on the calendar, dossier attached, CRM updated (post-approval), handoff acknowledged Skills Skill Type What it does Sequence writing Prompt 3 touches per committee role per channel, grounded in KB — no source, no claim Dossier personalisation Prompt Opens with the account's own trigger and hook, never a template smell Reply classification Logic Need × Fit × Authority × Timing × Intent state machine; misfits get a polite close

-- 5 of 8 --

Skill Type What it does Booking & warm handoff Logic Calendar invite + dossier to the human owner; unacknowledged → NORA's 36 h escalation Tools Warmed e-mail infrastructure Staggered sends, domain warm-up, caps and suppression built in LinkedIn messaging Connection + DM sequences within platform limits WhatsApp Business API Opt-in conversational touches where the ICP lives on WhatsApp Approval queue Every draft waits here — Approve / Edit / Skip; the gate at every level and every price Calendar & scheduler Booking against the human owner's availability CRM activity logger Every touch, reply and decision logged (writes gated) Guardrail: The approval gate is non-negotiable architecture. Higher autonomy levels widen batch approvals; they never remove the gate. 8 · ANALYST — analysis report (Sub-Agent 5) Closes the loop on the right side of the whiteboard. ANALYST reads the shared state, computes the funnel per ICP and per channel, and writes the report a human actually reads — what worked, what to change, shipped every Friday. Input Campaign state (Smriti) + CRM activity Output Weekly analysis report: funnel per ICP & channel, reply and refusal counts, next-run recommendations Definition of done Posted to the workspace channel Friday; numbers reconcile with the CRM Skills Skill Type What it does Funnel metrics Logic 100→80→15→75→meetings computed per ICP and per channel, deterministic from state Insight narration Prompt Explains movements and anomalies; proposes checklist changes for the next run Friday report Logic Sourced, drafted, sent, refused, booked — the department's weekly numbers, on schedule Tools State reader Read-only over Smriti and CRM activity Metrics engine Funnel, conversion and pacing computation Report generator Doc / dashboard output; feeds the console's metrics tiles Slack / Teams post Delivery of the Friday report Guardrail: Recommendations route back to PLAN as proposed checklist edits — humans approve strategy changes, same as sends.

-- 6 of 8 --

9 · ONBOARD — customer onboarding (Sub-Agent 6) The bottom-right box on the whiteboard, for when a deal lands: account access → customer calls → onboarding. ONBOARD activates the new customer and feeds everything it learns back into PLAN's Knowledge Base, so the next campaign starts smarter. Input A closed customer (from the human seller's handoff) Output Activated account: accesses provisioned, first campaign brief captured, day 0–7 plan running Definition of done Customer's first campaign live inside 7 days; context written back to KB Skills Skill Type What it does Access checklist Logic CRM connect, e-mail auth, calendar, channel accounts — tracked to completion Customer calls Prompt Schedules onboarding calls, summarises transcripts into structured context Day 0–7 activation plan Prompt Connect (Day 0) → target & knowledge (Day 1) → silent research (Days 1–2) → first 20–50 approvals (Day 3) → sends & replies (Days 4–6) → first qualified meeting (typically Day 7) Tools Access provisioning Workspace, integration and permission setup Calendar & meetings Onboarding call scheduling Transcript summariser Calls into structured, sourced KB entries CRM + KB writer Customer context flows back to PLAN (writes gated) 10 · Shared services — what every sub-agent stands on Smriti (shared memory) One workspace state across all six sub-agents: accounts, stages, decisions, learnings. No agent has private state that matters. Approval queue Single gate for all outbound and all CRM writes, regardless of which sub-agent produced the item. Knowledge Base PLAN writes it; everyone reads it; ENGAGE may only claim what it contains. Stale entries flagged by refresh rules. Audit log Every action by every sub-agent, customer-readable. Karma and level promotions are computed from this record. KARMIC (learning) Approval edits, refusals and outcomes feed the learning loop — what humans change is what the system learns. Deep research + web search The two base tools from the whiteboard, available in every sub-agent. 11 · Build order — from today's NORA to this architecture Phase 1 (now): PLAN, DISCOVER, ENRICH and ENGAGE already exist as behaviours inside NORA — split them into explicit sub-agents behind the same console, no user-visible change. Phase 2: ANALYST as a separate agent

-- 7 of 8 --

producing the Friday report and next-run recommendations. Phase 3: ONBOARD, triggered by closed-won — it is also the internal dogfood for the client self-serve onboarding deck. WhatsApp as ENGAGE's third channel lands with Phase 2, behind the same approval gate. Version 1.0 · 25 Aug 2026 · Sub-agent names (PLAN/DISCOVER/ENRICH/ENGAGE/ANALYST/ONBOARD) are internal working names — they are NORA's skills from the user's point of view, not separate products. SERA, DEXA and RIVA remain separate org-level agents and are unaffected by this split.

-- 8 of 8 --