Sales Development Agent

What's the smallest move that gets a stranger to say "tell me more"?

The Job​

When the kill date is 28 days away and there are zero paying customers, help the founder find qualified prospects, score them, compose personalized outreach, and sequence follow-ups — so human time goes to discovery calls, not research and admin.

The job: "I need 5 qualified conversations this month with construction/solar teams who have active RFP volume and spreadsheet pain."

The hidden objection: "AI outreach feels spammy." The agent must compose messages that a human would send — not templates with merge fields. The human reviews every outreach before send. The agent handles research, scoring, drafting, and timing. The human handles judgment, relationship, and trust.

Why This, Why Now​

28 days to kill date
0 paying customers
9 algorithms x 0 UI = 0 value
CRM built, RFP built, email built, LinkedIn built
Nobody is using any of it to generate demand

The product is ready. The demand generation isn't. This agent composes existing infrastructure into a demand generation machine.

The dogfooding principle: We are building a sales and marketing tool. If we can't use our own system to sell itself, the product doesn't work as advertised. Track 3 (founder outreach using the CRM) starts immediately — every friction point found is a bug, every smooth interaction is a proof point. The outreach IS the commissioning test.

Components Used​

This agent is one instrument in the Sales Process Optimisation orchestra. See SPO's composition inventory for the full platform stack.

Build ratio: ~80% composition, ~20% new code (new code is a subset of SPO's 25%).

Demand-Side Jobs​

Job 1: Find and Score Prospects Worth Talking To​

Situation: A founder knows their ICP — construction and solar EPC teams with active RFP volume. But finding specific companies requires hours of browsing, and once found, there's no system to rank which 5 of 20 deserve a call this week. The agent finds AND scores in one workflow — that's what the WorkChart shows.

Job 2: Say Something Worth Reading​

Situation: A founder drafts a cold outreach email. It's either too generic ("I'd love to explore synergies") or too salesy ("We can save you 83% on RFP time"). Neither gets a reply.

Job 3: Follow Up Without Forgetting​

Situation: A prospect opened the email but didn't reply. The founder meant to follow up on Wednesday. It's now Friday. The window is closing.

Job 4: Learn What Converts​

Situation: After a month of outreach, the founder has no idea which messages worked, which channels converted, or which ICP segment responded. Every week starts from zero.

The Agent Architecture​

This agent follows the Sales Work Chart's AI Transformation Spectrum. The agent handles AI-Led and AI-Only jobs. The human handles Human-Led and Human-Only jobs.

AGENT HANDLES (AI-Led + AI-Only)          HUMAN HANDLES (Human-Led + Human-Only)
─────────────────────────────────          ──────────────────────────────────────
Prospect research                          Discovery calls
Lead scoring + ranking                     Relationship building
Outreach drafting                          Message review + personalization
Follow-up sequencing                       Deal negotiation
Activity logging                           Account strategy
Channel optimization                       Trust building
CRM data hygiene                           Go/No-Go decisions
Email/LinkedIn scheduling                  Executive selling
Engagement tracking                        Pilot onboarding
Pipeline reporting                         Pricing conversations

WorkChart: Outreach Sequencing​

New WorkChart following the RFP workflow pattern. Five stages, adaptive routing by lead score.

STAGE 1: RESEARCH              STAGE 2: SCORE              STAGE 3: COMPOSE
Prospect Intelligence     →    Lead Scoring           →    Outreach Drafting
  Input: Company + ICP           Input: Research data        Input: Score + research
  Uses: DocumentIntel            Uses: sales-forecasting     Uses: SPCL scoring
  Output: Prospect profile       + explore-exploit           Output: Draft message
                                 Output: Fit score + rec     (email or LinkedIn)

STAGE 4: SEQUENCE              STAGE 5: MEASURE
Follow-up Cadence         →    Feedback Loop
  Input: Draft + score           Input: Send/open/reply data
  Uses: CRM tasks + email        Uses: compound-rate-tracking
  + LinkedIn adapters             + explore-exploit
  Output: Scheduled sequence     Output: Channel performance
  (human approves before send)   + message effectiveness

Routing Logic (Adaptive)​

Agent Profile — Phygital Being​

The Sales Dev Agent is a phygital being — same pattern as Truth Terminal, AIXBT, Luna. Public identity lives in the dream repo at /docs/players/phygital-beings/. Engineering loads the profile via the Agent Platform ETL CLI.

Dependency: Agent Continuity PRD Sprint 0 — add --path flag to agent-etl-cli.ts so new agents can load from any filesystem location. This is a ~20 line change that unblocks all future agent profiles.

Profile location (dream repo): docs/players/phygital-beings/sales-dev-agent/

Profile data (8 JSON files following existing convention):

The docs-as-context loop: The phygital being's documentation in /docs/players/phygital-beings/ serves three audiences simultaneously:

1. External — Visitors to mm.dreamineering.com see how AI agents work in the system
2. Internal agents — Engineering team agents read the profile spec as build context
3. The agent itself — Semantic and procedural memory loaded from these docs at session start

This reinforces quality of docs here and sells ideas to humans and agents alike.

A2A Handoff​

When a deal enters Proposal stage in the CRM pipeline, the Sales Dev Agent hands off to the RFP Agent via A2A protocol:

Sales Dev Agent                          RFP Agent
───────────────                          ─────────
Prospect → Qualify → Outreach →          → Proposal → Review → Submit
  Discovery call → Deal created   [A2A]    RFP upload → AI auto-fill
                                           → Answer generation
                                           → Deck composition

Feature / Function / Outcome​

Delivery Summary​

Success Criteria​

Functional​

Outcome (Kill Date: 2026-03-24)​

Team-level outcomes (discovery calls booked, pipeline value, first paying customer) are tracked in Sales Process Optimisation.

Business Dev​

Commissioning​

Risks + Kill Signal​

Kill signal: If 50 outreach messages produce 0 replies across both email and LinkedIn over 30 days, the ICP or the message is wrong. Diagnose with explore-exploit data before iterating.

Build Sequence​

Sprint -1: Prove Outbound Infrastructure (3 days engineering)​

Before building the agent, prove the delivery channels work end-to-end. This is the foundation everything else depends on.

Email (Resend) — Go Live (0.5 day)​

The full code path exists: UI → sendEmailAction → SendEmailUseCase → ResendAdapter → Resend API → webhook → CRM activity update. Nothing to build. Configuration only.

Engineering tasks:

1. Set environment variables in Vercel:

   • RESEND_API_KEY — from Resend dashboard
   • RESEND_FROM_EMAIL — verified domain email (e.g. outreach@dreamineering.com)
   • RESEND_WEBHOOK_SECRET — generate in Resend webhook settings

2. Configure Resend webhook endpoint:

   • URL: https://dreamineering.com/api/webhooks/resend
   • Events: email.delivered, email.opened, email.clicked, email.bounced

3. Send a test email from Settings > Marketing compose modal

4. Verify webhook fires back — check CRM activity customProperties.emailEvents has delivery event

Acceptance test: Send email → recipient opens it → webhook fires → CRM activity shows "email.opened" event with timestamp.

LinkedIn — Replace Mock Client with Real API (2.5 days)​

The adapter is 1043 lines. ~990 lines are real (validation, metrics, composition, B2B keyword detection, content rules, server actions, UI). The mock is ~50 lines in initializeClient() that return synthetic IDs instead of making API calls. Replace the mock, keep everything else.

Engineering tasks:

1. Register LinkedIn app (30 mins + approval wait)

   • LinkedIn Developer Portal → Create App
   • Products: "Share on LinkedIn" + "Sign In with LinkedIn using OpenID Connect"
   • Note: LinkedIn app review can take days — start this immediately, parallel with code work

2. Implement OAuth 2.0 token exchange (1 day)

   • File: New linkedin-oauth.ts in libs/infrastructure/adapters/src/social-media/
   • 3-legged OAuth flow: redirect → callback → token exchange
   • Store accessToken + refreshToken + expiresAt (env vars or encrypted DB)
   • Token refresh on 401 response
   • Reference: LinkedIn Marketing API v2 auth docs

3. Replace mock initializeClient() (1 day)

   • File: libs/infrastructure/adapters/src/social-media/linkedin.adapter.ts lines ~860-914
   • Replace synthetic returns with real HTTP calls to LinkedIn REST API v2:
     • posts.create → POST https://api.linkedin.com/v2/posts (using author URN from OAuth profile)
     • posts.delete → DELETE https://api.linkedin.com/v2/posts/{id}
     • posts.get → GET https://api.linkedin.com/v2/posts/{id}
     • analytics.getPostStats → GET https://api.linkedin.com/v2/organizationalEntityShareStatistics
     • profile.getMe → GET https://api.linkedin.com/v2/userinfo (OpenID Connect)
   • The adapter already handles error mapping, metrics, resilience — just wire the HTTP client

4. Set environment variables:

   • LINKEDIN_CLIENT_ID — from Developer Portal
   • LINKEDIN_CLIENT_SECRET — from Developer Portal
   • LINKEDIN_ACCESS_TOKEN — from OAuth flow
   • LINKEDIN_REFRESH_TOKEN — from OAuth flow

5. Test from UI: Settings > Marketing > LinkedIn Post Composer → compose post → click Post → verify it appears on actual LinkedIn profile

Acceptance test: Compose post in app → post appears on real LinkedIn → adapter returns real postId → getPostStatus() returns published status.

Wire Both to CRM Activity Logging (0.5 day)​

Currently email webhooks update CRM activities but the initial send doesn't create one. LinkedIn has no activity logging at all.

1. When sendEmailAction succeeds: create CRM activity (type: "email", linkedContactId, linkedDealId, description: subject + preview, customProperties: { emailMessageId, channel: "email" })
2. When postToLinkedInAction succeeds: create CRM activity (type: "note", description: post text, customProperties: { linkedinPostId, channel: "linkedin" })

This closes the loop: every outbound touchpoint appears in the CRM activity timeline automatically.

Sprint -1 Commissioning​

Track 1: Wire algorithms to CRM UI (engineering — 6 days) This is the existing Tier 1 sprint from the Sales CRM PRD. Sales Forecasting, Compound Rate, RFP Type Detection, Stripe. Makes the product demo-ready. Runs in parallel with Sprint -1.

Track 2: Sales Dev Agent (engineering — 14 days)

Track 3: Dogfood outreach (founder — starts immediately) Don't wait for the agent. Start discovery calls now using the CRM, email, and LinkedIn tools we've built. The outreach itself IS the product demo — if we can't use our own system to generate demand, why would anyone else?

This is the commissioning test for the Sales CRM before any customer sees it:

Every outreach touchpoint generates data that proves the system. Every friction point discovered is a bug to fix before pilots. The founder's dogfood loop IS the commissioning protocol for Jobs 1-5.

The Tight Five for this outreach:

1. Principles — "Win more bids" is the value proposition, grounded in compounding answer libraries
2. Performance — Reply rate, meeting conversion, pipeline value generated
3. Platform — CRM + RFP + Email + LinkedIn, all wired
4. Protocols — Outreach sequence: research → score → compose → send → follow up → log
5. Players — Construction/solar EPC teams with 10+ employees and active RFP volume

Relationship to Other PRDs​

This PRD is Horizon 1 of a three-horizon system:

HORIZON 1: PROVE IT WORKS (Now → 90 days)            ← THIS PRD
  Sales Dev Agent — one instrument
        │ feeds into
        ▼
HORIZON 2: ORCHESTRATE THE TEAM (90 days → 6 months)
  Sales Process Optimisation — the orchestra
        │ generates data for
        ▼
HORIZON 3: MAKE IT A PROTOCOL (6 months → 18 months)
  Trust Commerce — the protocol

Data flows:

• This agent → SPO: Outreach data, reply patterns, and qualification outcomes feed SPO's Learning Agent to improve ICP scoring across all agents.
• SPO → this agent: 4-layer ICP scores, workflow dispatch, optimized routing rules.
• Trust Commerce → this agent: Once live, warm referral leads arrive pre-qualified by the trust graph, changing scoring and outreach approach. Cold outreach becomes unnecessary for the portion of pipeline that comes through the trust network.

Dependencies:

• Sales Process Optimisation — Composition inventory, ICP validation model, team-level success criteria
• Sales CRM & RFP — Sprint 0 algorithm wiring must complete first
• Agent Platform — ETL CLI --path flag for profile loading

Mycelium Capability​

The Sales Development Agent is a demand generation capability. It composes CRM, email, social, and intelligence algorithms into an outreach pipeline. Currently dogfooded on construction and solar ICP, and deployed externally for Sneakers Media's HRV campaign — the same agent, two contexts. This is the BOaaS proof: one instrument serving multiple ventures.

The WorkChart pattern (Research → Score → Compose → Sequence → Measure) applies to any B2B outbound motion.

Currently Growing In: stackmates

Context​

• Sales CRM & RFP — The convert/retain half this agent feeds
• Sales Work Chart — Human/AI split that defines the agent's boundaries
• Social Media Marketing — LinkedIn compose-schedule-learn loop
• Content Amplifier — Multi-platform distribution
• ICP Framework — Target definition
• Jobs To Be Done — Demand-side thinking
• Phygital Mycelium — The capability catalogue
• Commissioning Dashboard — Progress tracking