Research Agent → Copy Agent → QA Agent: The Multi-Agent Outbound Workflow That Doesn’t Burn Your Domain

Most outbound teams don’t “scale outbound.” They scale mistakes.

They stack five tools, duct-tape them with Zapier, spray 10,000 prospects, then act surprised when deliverability tanks and reps drown in junk replies. Your domain didn’t “randomly” burn. Your workflow did.

A multi-agent outbound workflow fixes that. Not because “AI.” Because assembly lines beat chaos.

TL;DR

• Run outbound like a factory: Research → Enrich → Copy → QA → Sequence → Reply.
• QA is the profit center. It prevents: bad lists, risky claims, spammy phrasing, duplicates, and sending to the wrong people.
• Your stop rules protect deliverability more than “warming” ever will.
• Mailbox providers now enforce stricter authentication and bulk-sender standards. Outlook specifically announced rejection for high-volume senders that fail SPF, DKIM, and DMARC starting May 5, 2025. (Microsoft Tech Community)
• Inbox placement is not “delivered.” It’s “inbox vs spam vs missing.” Validity’s benchmarks show deliverability is still fragile. (Validity)
• Chronic runs the whole thing end-to-end, till the meeting is booked. No tool soup.

---

Define the multi-agent outbound workflow (so you can actually build one)

Multi-agent outbound workflow: a coordinated set of specialized agents that execute outbound in stages, where each agent has:

• a single job,
• strict inputs and outputs,
• hard QA gates,
• stop rules that prevent deliverability and brand damage.

Think: assembly line, not “AI writes emails.”

The agent assembly line (blueprint)

1. Research Agent - confirms ICP match and finds triggers
2. Enrichment Agent - finds the right person, verifies contactability, adds technographics
3. Copy Agent - writes a tight personalized opener and one clear offer
4. QA Agent - blocks bad data, risky claims, spamminess, duplicates, and compliance issues
5. Sequencing Agent - builds cadence, branching logic, and fallback channels
6. Reply Agent - classifies replies and routes to the right next action fast

This is the workflow that doesn’t burn your domain because it stops garbage before it ships.

---

Why domains burn (it’s not “bad luck”)

Domains burn for three boring reasons:

1. Bad targeting
   • Wrong ICP = more deletes, more spam complaints, more “stop emailing me.”
2. Bad data
   • Bounces, dead inboxes, role accounts, catch-alls you never validated.
3. Bad behavior
   • Over-emailing, spammy copy, no opt-out, ignoring suppression, sending duplicates.

Mailbox providers reward patterns. They also punish patterns.

And they got louder about it. Outlook announced new requirements for high-volume senders and enforcement that can include rejecting non-compliant email if SPF, DKIM, and DMARC are not set up correctly. The enforcement date they stated was May 5, 2025. (Microsoft Tech Community)

If your “workflow” ignores basics like authentication and list hygiene, you are not running outbound. You are running a slow-motion blocklist speedrun.

---

Agent 1: Research Agent (ICP match + triggers)

Job

Pick the right accounts. Find a reason now.

Inputs

• ICP definition (firmographics + negatives)
• Buying signals to watch
• Regions, languages, compliance constraints
• Your offer types (what you can credibly sell)

If your ICP lives in someone’s head, this agent fails. Put it in writing.

Use an ICP builder or die by vibes. Chronic has one: ICP Builder

Outputs (what the agent must produce every time)

For each account:

• ICP Match: yes/no + score + why
• Trigger(s): 1-3 specific events with timestamps
• Angle: which offer maps to which trigger
• Do-not-target flags: competitors, recent layoffs, regulated segment you can’t touch, etc.

Trigger menu (steal this)

Pick triggers your buyers actually react to:

Hiring and org changes

• “Hiring SDRs” or “Hiring RevOps”
• New VP Sales, new Head of Growth

Tech stack

• New CRM, new data tool, new marketing automation
• Tool churn (ripping out Salesforce, moving to HubSpot)

Timing events

• Funding
• New product launch
• New market expansion
• New location

Operational pain

• Job posts that scream “manual work”
• Website performance issues, broken funnel steps
• “We’re scaling outbound” in a podcast post (yes, people say this out loud)

Research Agent template (output schema)

Use this exact structure so downstream agents don’t guess.

Account Research Card

• Account: {{company_name}}
• URL: {{company_domain}}
• ICP Score (0-100): {{icp_score}}
• ICP Reasons:
  • {{reason_1}}
  • {{reason_2}}
• Triggers (with proof):
  • Trigger 1: {{trigger_1}} - Source: {{url}} - Date: {{date}}
  • Trigger 2: {{trigger_2}} - Source: {{url}} - Date: {{date}}
• Primary angle: {{angle}}
• Exclusions: {{exclusion_flags}}
• Notes for copy: {{copy_notes_1_line}}

---

Agent 2: Enrichment Agent (contacts + technographics)

Job

Turn “good account” into “right person with reachable contact.”

Inputs

• Account Research Card
• Target personas (DM list)
• Seniority band (manager? director? VP?)

Outputs

• 1-2 primary contacts + 1 backup contact
• Verified email + phone when available
• Role relevance score
• Technographics that matter to your offer
• Suppression checks

Chronic bakes enrichment in: Lead enrichment

Technographics that actually matter

Don’t collect trivia. Collect ammunition.

Good:

• CRM (Salesforce, HubSpot, etc.)
• Sales engagement (Outreach, Salesloft)
• Data providers (Apollo, ZoomInfo)
• Website platform (Shopify, Webflow) when relevant
• Analytics and intent tooling if it changes your pitch

Bad:

• “They use React.” Nobody is buying meetings because you detected React.

Enrichment Agent template (output schema)

Contact Enrichment Card

• Account: {{company_name}}
• Primary contact:
  • Name: {{first_name}} {{last_name}}
  • Title: {{title}}
  • LinkedIn: {{linkedin_url}}
  • Email: {{email}} (status: {{verified|risky|unknown}})
  • Phone: {{phone}} (status: {{direct|main|unknown}})
• Backup contact:
  • Name: {{...}}
• Persona fit: {{persona_fit_score}}/100
• Technographics:
  • CRM: {{crm}}
  • Email/Sales tool: {{sequencer}}
  • Data provider: {{data_tool}}
• Suppressions:
  • Prior contacted? {{yes/no}}
  • Unsubscribed? {{yes/no}}
  • Competitor/customer? {{yes/no}}
  • Role account? {{yes/no}}

---

Agent 3: Copy Agent (personalized opener + offer)

Job

Write one email that earns a reply.

Not a novel. Not a brochure. Not a “quick question” drive-by that wastes everyone’s time.

Inputs

• Account Research Card
• Contact Enrichment Card
• Offer library (2-3 offers max)

Chronic writes personalized copy inside the workflow: AI email writer

Output rules (non-negotiable)

• Personalization stays in the first 1-2 lines.
• One outcome. One ask.
• No fake familiarity. No “loved your post” unless you actually reference the post.
• No inflated claims. If you can’t prove it, cut it.

Personalization fields (copy variables template)

This is the minimum set that produces real personalization without writing fan fiction:

Account-level fields

• {{company_name}}
• {{industry}}
• {{employee_count_range}}
• {{location}}
• {{trigger_1}}
• {{trigger_1_date}}
• {{trigger_proof_url}}
• {{current_tool}} (if relevant)

Contact-level fields

• {{first_name}}
• {{title}}
• {{function}} (RevOps, Sales, Growth)
• {{persona_pain_1}} (from your playbook, not guessing)

Offer-level fields

• {{outcome_metric}} (meetings/month, time saved/week)
• {{offer_type}} (audit, teardown, pilot, benchmark)
• {{cta}} (15 min, 2 questions, etc.)

Copy templates (email 1 options)

Template A: Trigger → outcome → question (tight)

Subject: {{company_name}} + outbound

Hi {{first_name}} - saw {{trigger_1}} ({{trigger_1_date}}).

When {{persona_pain_1}} shows up right after that, pipeline gets noisy fast.

Worth a 15-min look at your outbound workflow and where it’s leaking replies and meetings?

• {{sender_name}}

Template B: Tool stack angle (only if you have proof)

Subject: quick question

Hi {{first_name}} - noticed {{company_name}} uses {{current_tool}}.

Teams usually hit the same wall: research in one place, enrichment in another, sequences somewhere else, then nobody trusts the CRM.

Want me to send a 6-line teardown of your current outbound workflow with 3 fixes?

• {{sender_name}}

Template C: “Two bullets” (for operators)

Subject: {{company_name}} outbound question

{{first_name}} - quick one.

If you’re pushing outbound this quarter, what’s the bigger bottleneck right now?

1. list accuracy, or
2. reply handling speed

I can share the exact workflow we use to fix whichever one you pick.

• {{sender_name}}

---

Agent 4: QA Agent (claims, compliance, spamminess, duplication)

Job

Block everything that kills deliverability or trust.

QA is where most teams cheap out. Then they pay for it with a dead domain.

Inputs

• Draft email
• Research + enrichment cards
• Suppression lists
• Claim library (what you can say, what you cannot)
• Compliance rules (region, opt-out, internal policies)

Outputs

• Pass/fail + reasons
• Corrected draft if fixable
• Hard stop if not

The QA checklist (copy/paste and use)

Use this as a gate. If it fails, it doesn’t send.

A) List and identity checks

• Recipient email status is verified or approved-risk (defined)
• Not a role account (info@, sales@, support@) unless explicitly allowed
• Not on suppression list (unsubscribe, do-not-contact, competitor, customer)
• Not contacted in last {{X}} days (dedupe window)
• Account meets ICP score threshold (ex: 70+)

B) Claim integrity checks

• Every numeric claim has a source or internal proof
• No fake personalization (no “saw your post” with no link)
• No implied partnership with tools or brands without approval
• No sensitive inferences (health, finances, private data)

C) Compliance and policy checks

• Includes opt-out line when required by your policy
• Sender identity is real (name, company, reply-to works)
• No misleading subject lines

D) Deliverability and “spamminess” checks

• Under 120 words (default)
• One link max (optional, none is fine)
• No image, no tracking pixel (optional policy)
• Avoids classic spam formatting: all caps, multiple exclamation points, weird spacing
• Doesn’t stack “free”, “guarantee”, “act now” style phrasing
• Subject line is plain and human

Mailbox providers also tightened expectations around complaint rates and deliverability hygiene. Validity’s reporting frames deliverability as inbox placement, not “sent minus bounces,” and highlights how much mail never reaches the inbox. (Validity)

E) Duplication checks (quiet domain killer)

• Opener is not duplicated across >{{N}} prospects in a batch
• Body similarity score under {{threshold}}
• No repeated sentence templates across an entire segment

QA Agent “stop the line” rules

If any of these happen, pause the campaign automatically:

• Bounce rate > 2% in last 200 sends
• Spam complaints exceed provider thresholds (track by provider)
• “Not interested” plus “unsubscribe” replies spike above baseline
• Open-to-reply ratio collapses (sign of inboxing issues or mismatched targeting)
• More than {{X}} duplicate openers detected

Stop rules beat panic later.

---

Agent 5: Sequencing Agent (cadence + branching)

Job

Build a sequence that respects attention and adapts to signals.

Inputs

• QA-approved email 1
• Channel options (email only, email + LinkedIn, etc.)
• Risk tolerance (domain age, volume, segment)

Outputs

• Sequence steps
• Wait times
• Branching logic
• Stop rules per prospect

Sequence skeleton (safe default)

This is a template_post, so here’s a sequence you can ship today.

Sequence: 10 business days, 4 touches

1. Day 1 (Email 1) - trigger-based opener, one question
2. Day 3 (Email 2) - proof or artifact (teardown, benchmark, quick loom if you do that)
3. Day 6 (Email 3) - objection flip, clarify who it’s for and not for
4. Day 10 (Email 4) - clean breakup, confirm routing (“wrong person?”)

Branching logic (what most sequences miss)

Branch on what the prospect does.

• If positive reply: route to booking flow immediately
• If neutral reply (“maybe later”): drop into nurture, set a task, stop sequence
• If objection: send the objection-specific snippet, stop other touches
• If unsubscribe/remove: suppress instantly, no exceptions
• If out-of-office: pause until return date, then resume with a soft restart

Sequencing “stop rules” template (per prospect)

• Stop sequence if:
  • Reply received (any)
  • Unsubscribe received
  • Hard bounce
  • Manual do-not-contact flag added
  • Meeting booked
• Pause sequence if:
  • OOO detected with return date
  • Prospect changes job (trigger re-research)

---

Agent 6: Reply Agent (classification + routing)

Job

Replies are the product. Route them fast.

Slow response time kills meetings. Also, it wastes the one moment your prospect is paying attention.

Inputs

• Incoming replies
• Account owner rules
• Calendar links and booking policy
• Qualification rules (fit + intent)

Outputs

• Reply classification
• Next action
• Owner assignment
• Draft response (optional) that matches your policy

Reply categories (minimum viable taxonomy)

• Positive (interested, asks for time)
• Clarifying question (pricing, scope, “what do you do”)
• Referral (“talk to Alex”)
• Objection (timing, budget, already have vendor)
• Unsubscribe / do-not-contact
• Negative
• OOO
• Spam complaint risk (hostile responses, “reporting you”)

Routing rules (simple and ruthless)

• Positive, referral, clarifying: notify owner instantly and draft response
• Objection: route to objection playbook response
• Unsubscribe: suppress immediately and confirm politely once
• OOO: pause sequence and schedule re-touch
• Hostile: suppress and do not reply more than once

Chronic’s core point here is system behavior. You want pipeline on autopilot, not “inbox triage on autopilot.” That’s why the workflow ends at the meeting being booked, not “email sent.”

---

Put the workflow in one system (or enjoy your tool spaghetti)

Most teams run this “multi-agent outbound workflow” across:

• a lead database,
• an enrichment tool,
• a personalization tool,
• a sequence sender,
• a CRM,
• a reply classifier,
• plus Zapier babysitting the whole thing.

That’s how you get:

• duplicates,
• missing fields,
• leads stuck in limbo,
• reps emailing suppressed contacts,
• and “why did we send this” moments.

Chronic runs the full assembly line:

• ICP definition with ICP Builder
• Data and contacts with Lead enrichment
• Prioritization with AI lead scoring
• Copy with AI email writer
• Workflow visibility in the sales pipeline

End-to-end, till the meeting is booked.

And if you’re comparing stacks:

• Chronic vs Apollo: Chronic vs Apollo
• Chronic vs HubSpot: Chronic vs HubSpot
• Chronic vs Salesforce: Chronic vs Salesforce

One line of contrast: Clay is powerful and complex. Instantly sends emails. Salesforce costs a fortune and still needs bolt-ons. Chronic runs the workflow.

---

Template pack: copy/paste assets (the “template_post” part)

1) Personalization field map (minimum viable)

Use this table to standardize fields across agents.

2) QA checklist (operational version)

Paste into Notion. Turn into a form. Make it painful to skip.

QA Gate: PASS requires all

• Identity and suppression: PASS/FAIL
• ICP threshold: PASS/FAIL
• Trigger proof exists: PASS/FAIL
• Claims: PASS/FAIL
• Spamminess: PASS/FAIL
• Duplication: PASS/FAIL
• Sequence stop rules set: PASS/FAIL

If FAIL

• Fixable? (Y/N)
• Fix notes:
• Re-run QA after fix

3) Sequence skeleton (with branching)

Sequence Name: Triggered outbound (4-touch)

Email 1 (Day 1): Trigger mention + one question
Email 2 (Day 3): Share artifact (teardown/benchmark)
Email 3 (Day 6): Objection preemption + narrow ICP
Email 4 (Day 10): Breakup + wrong person check

Branching

• If reply: stop all, route to Reply Agent
• If OOO: pause, resume on return
• If click but no reply: send Email 2 variant with artifact
• If no open after Email 2: tighten subject, shorten Email 3

4) Stop rules (campaign-level)

Stop the entire campaign if:

• Hard bounce rate > 2% over last 200 sends
• Unsubscribe rate spikes above baseline by 2x
• Complaint risk signals rise (hostile replies, block threats)
• Duplicate opener detection > {{N}} in a batch

---

Two posts worth reading before you blame deliverability gods

• Cold Email Deliverability in 2026: Relevance Beats DNS Now (Here’s the Fix)
• Dual Scoring Template: Fit + Intent, With the Minimum Signals That Actually Work
• The New Outbound Stack in 2026: Why “One More Tool” Kills Pipeline

---

FAQ

What’s the real point of a multi-agent outbound workflow?

Consistency. Each agent produces a clean output that the next agent can trust. That prevents bad lists, bad claims, and bad sequencing from hitting your domain at scale.

Which agent matters most for deliverability?

The QA Agent. It blocks bounces, duplicates, suppression misses, and spammy copy before anything sends. It also enforces stop rules when metrics turn.

How many triggers should the Research Agent use per prospect?

One is enough. Two max. If you need three triggers to justify emailing someone, you’re stretching. That reads like stalking, not relevance.

What compliance changes matter most right now for outbound?

Authentication and sender standards keep tightening. Outlook announced rejection for high-volume senders that do not meet SPF, DKIM, and DMARC requirements starting May 5, 2025. (Microsoft Tech Community)

What benchmarks should we expect from cold email in 2026?

Mid-single-digit reply rates are common in large benchmark discussions. Belkins published analysis on cold email response rates based on millions of emails, and their content highlights variability by structure and approach. Use benchmarks as guardrails, not goals. (Belkins)

How do we run this without five tools and Zapier?

Use one system that owns the whole flow: ICP → enrichment → scoring → copy → QA gates → sequencing → reply routing → pipeline. Chronic is built for that, including AI lead scoring, lead enrichment, and a real sales pipeline.

---

Build the line, then hit send

Pick one ICP. One trigger set. One offer.

Then implement the assembly line:

1. Research card
2. Enrichment card
3. Copy draft
4. QA gate
5. Sequence with stop rules
6. Reply routing

If your outbound can’t pass QA at 50 prospects, it has no business touching 5,000. That’s how domains die.