Our AI System

Your pipeline runs on 15 agents built on Claude.

We didn't wire together a stack of SaaS tools and call it AI. We built a multi-agent system on Anthropic's Claude that researches accounts, writes content, monitors deliverability, and learns — improving every campaign with everything it's seen before.

Book a CallSee the System
Orchestrator

Claude Opus 4

Discovery
1 agent
Targeting
4 agents
Content
3 agents
Execution
3 agents
Meeting
2 agents
Inbound
1 agent
Learning
The compounding moat
2 agents · always-on
Project Memory
per client
Cross-Client Repository
all clients
15
Specialized AI agents
16
MCP tool integrations
2.4×
Better campaigns from learning loops
Day 1
New clients benefit from prior learning
How the System is Built

One orchestrator. Six clusters. 15 agents.

Claude Opus 4 sits at the top, routing every task to the right specialist agent. Each agent has a single job, its own tools, and its own memory access. Adding a new capability means adding a new agent — not rebuilding something that already works.

Orchestrator

Claude Opus 4

Routes every task to the right agent · judges, synthesizes, and delegates

01 · Discovery
1 agent

Finding your best buyers

Discovery Agent
02 · Targeting
4 agents

Scoring & researching accounts

Research AgentAccount ScoringSignal DetectionStakeholder Mapping
03 · Content
3 agents

Writing and quality control

Content AgentQA AgentCompliance Agent
04 · Execution
2 agents

Sending and reply handling

Reply TriageDeliverability Monitor
05 · Meeting
2 agents

Post-meeting intelligence & momentum

Meeting IntelligenceDeal Momentum
06 · Inbound
1 agent

Qualifying inbound leads

Inbound Triage
07 · Learning · The Compounding Moat

Patterns that improve every campaign

2 agents
Targeting Feedback AgentRepository Agent
Memory Architecture

Most agencies keep learning in people's heads.

When someone leaves, the memory leaves. Thyleads' memory lives in the system — per-client knowledge that compounds over every engagement, and cross-client patterns that every new client inherits from day one.

Project Memory

A persistent knowledge graph specific to each client. Every agent that runs on your campaign has read access to what's been learned — ICP briefs, customer interview transcripts, stakeholder maps, campaign results, meeting outcomes.

  • Kickoff brief & ICP brief (all agents read this)
  • Customer interview transcripts + language bank
  • Account universe with scores and signals
  • Won/lost deal analysis
  • Campaign variants + reply performance
  • Meeting outcomes, objections raised, competitors named

Cross-Client Repository

Anonymized patterns aggregated across all clients. A client onboarded in month 18 benefits from 17 months of learning — without anyone manually telling the system what works. This is what turns operations into a compounding competitive advantage.

  • Skeleton performance by ICP type and industry
  • Best send-time patterns by persona (day + hour)
  • Subject line patterns that consistently pull replies
  • Objection patterns by industry (e.g. FinTech vs HRTech)
  • Channel mix data — when WhatsApp lifts conversion
  • Patterns promoted only when seen across 3+ clients
How the Learning Works

A deal on Day 30 makes Day 90 campaigns better.

This is a real example of how one meeting outcome flows through the system and improves the next campaign — automatically, with no manual intervention.

1

Meeting happens

A VP Marketing at Razorpay mentions integration time vs Amplitude as a concern. Meeting Intelligence Agent extracts the objection, competitor named, and verbatim phrasing within 5 minutes of the call ending.

2

Pattern surfaces

Targeting Feedback Agent runs nightly. It sees the same objection in 6 of the last 12 meetings. It surfaces an insight: integration time vs Amplitude is a top-3 objection — currently not addressed in any first-touch variant.

3

Content Agent updates

A recommendation is pushed to Content Agent’s system prompt for the next generation run: “When targeting MarTech VP Marketing, address integration time relative to Amplitude. Use phrasing close to ‘while they take 4+ weeks, our standard is 7 days.’”

4

Reply rate improves 1.7×

New variants go out on Day 35 with the updated angle. Reply rate on the new variants is 1.7× the prior baseline. That pattern is stored. If 3+ clients show the same dynamic, it gets promoted to the Repository.

Agent Specifications

What each agent actually does.

15 agents. Each with a single job, a model tier matched to the task, and a defined learning loop. Haiku handles high-volume classification cheaply. Opus handles judgment, generation, and synthesis.

Discovery Agent

Opus 4 · Tier 1

Helps GTM engineers run better client discovery. Suggests category-specific interview questions, transcribes and structures customer conversations, and extracts the language bank — the verbatim phrases from real customer interviews that show up in replies later.

Structured interview summaryLanguage bankDraft ICP brief
Tool Layer

16 MCP integrations. One clean interface.

Agents access external tools through MCP servers — one per tool, one clean schema. Agents call MCPs, MCPs handle auth, rate limits, and response normalization. Your campaign never breaks because a rate limit was missed.

Prospecting

Apollo MCP

Contact and company database used by Research and Stakeholder agents.

Prospecting

Crustdata MCP

B2B data and hiring signals for Research and Signal Detection.

Inbound

Apify MCP

Web scraping across LinkedIn, news, job boards, and filings.

Enrichment

Cosy MCP

Enrichment orchestration with AI columns used by Research and Content.

Research

Sales Navigator MCP

Deep LinkedIn search for Research and Stakeholder Mapping agents.

Verification

ZeroBounce MCP

Email verification before any contact is added to a sequence.

Email Ops

Smartlead MCP

Campaign metrics and mailbox health for Deliverability and Reply Triage.

LinkedIn Ops

HeyReach MCP

LinkedIn campaign metrics plus cadence for Deliverability and Reply Triage.

Meetings

Fireflies MCP

Transcript ingestion for Meeting Intelligence Agent.

Scheduling

Google Calendar MCP

Meeting scheduling and event creation for Deal Momentum Agent.

Deliverability

Postmaster Tools MCP

Domain reputation and spam-rate monitoring.

Deliverability

MXToolbox MCP

Records and DNS checks for Deliverability Monitor Agent.

Outreach

Gupshup MCP

WhatsApp template sends for Deal Momentum Agent.

Verification

Gmail / Outlook MCP

1-to-1 email tools for AE assist after handover.

Sourcing

Hunter / LeadMagic MCP

Email-finding fallback for Stakeholder Mapping.

Memory

Thyleads Dashboard MCP

Read/write to Project Memory — used by all agents.

The Human Layer

Behind every agent, there's a human.

SStrategist
RResearcher
AAccount Manager

Every agent is supervised by our GTM pod: strategists, researchers, and account managers who shape the playbook, approve every message, and step in where AI can't judge context. The result is outbound that scales like a system and sounds like your best AE.

See how we work

Stop sending outbound that sounds like outbound.

See what a 15-agent system built on Claude can book in your calendar inside the first 30 days.

Book a Meeting