Multi-Agent Handoff Architect

Multi-Agent Handoff Blueprint

Executive Architecture

The proposed workflow should use a controlled sequential pipeline with conditional return paths rather than unrestricted agent-to-agent delegation.

Recommended architecture

Research
→ Research Validation
→ Strategy
→ Strategy Validation
→ Copy + Visual Prompt Production
→ QA
→ Human Approval
→ Client Delivery

The main architectural principle is

NO DOWNSTREAM AGENT SHOULD CONTINUE WHEN ITS REQUIRED INPUT CONTRACT IS INVALID.

This prevents weak research from contaminating every later stage.

---

Agent Map

Agent 01 — Research Agent

Responsibility

Collect and structure verified market, competitor, product, and customer information.

Can

• Research approved sources
• Extract facts
• Compare competitors
• Identify patterns
• Summarize evidence
• Assign confidence

Cannot

• Decide final campaign strategy
• Invent unsupported customer insights
• Write final campaign copy
• Approve its own research quality

Required Output

ResearchPackage

Must contain

• research_objective
• findings
• competitors
• customer_insights
• verified_claims
• source_references
• uncertainties
• confidence_score

Exit Gate

Proceed only if

confidence_score >= 0.75

and

all critical claims have supporting evidence.

If confidence < 0.75:

RETURN TO RESEARCH.

---

Handoff 01

Research Agent
→ Strategy Agent

Handoff Contract

Required fields

research_summary
verified_claims
competitor_findings
audience_findings
sources
uncertainties
confidence_score

Validation

The Strategy Agent must reject the handoff when:

• research confidence is below threshold
• critical source fields are missing
• unsupported claims are present
• target audience information is insufficient

The Strategy Agent must not repair missing research by inventing assumptions.

---

Agent 02 — Strategy Agent

Responsibility

Convert validated research into campaign strategy.

Can

• Define campaign objective
• Develop positioning
• Choose messaging angles
• Prioritize customer pain points
• Define content structure

Cannot

• Introduce unsupported market claims
• Modify verified research facts
• Generate final copy
• Approve its own strategy

Required Output

StrategyPackage

Contains

• campaign_objective
• target_audience
• positioning
• core_message
• proof_points
• creative_angles
• content_requirements
• prohibited_claims
• confidence_score

Exit Gate

Proceed when

strategy completeness = PASS

and

confidence_score >= 0.80

---

Handoff 02

Strategy Agent
→ Copywriting Agent
and
→ Visual Prompt Agent

This handoff may branch in parallel.

Both agents receive the same approved StrategyPackage.

They must not receive the entire raw research history unless required.

This reduces context overload.

---

Agent 03 — Copywriting Agent

Responsibility

Create campaign copy from approved strategy.

Can

• Write headlines
• Captions
• CTA variants
• Ad copy

Cannot

• Introduce new factual claims
• Change positioning
• Create unsupported statistics
• Override strategy constraints

Output Contract

CopyPackage

• headline_variants
• body_copy
• captions
• CTA
• claims_used
• strategy_reference

---

Agent 04 — Visual Prompt Agent

Responsibility

Translate strategy into production-ready visual prompts.

Can

• Define visual concept
• Camera direction
• Scene
• Lighting
• Composition
• Prompt structure

Cannot

• Alter factual product attributes
• Introduce unsupported product claims
• Rewrite campaign strategy

Output Contract

VisualPackage

• concept
• image_prompts
• video_prompts
• visual_constraints
• product_consistency_rules
• strategy_reference

---

Synchronization Gate

CopyPackage
+
VisualPackage
↓
QA Agent

QA must wait until both packages are complete.

Partial execution is not allowed.

---

Agent 05 — QA Agent

Responsibility

Validate consistency, quality, compliance, and cross-agent alignment.

QA Checks

• Copy matches strategy
• Visual direction matches strategy
• Claims trace to approved research
• No prohibited claims
• No contradiction between copy and visual
• Required deliverables are complete

QA Decision

PASS

REVISE_COPY

REVISE_VISUAL

REVISE_STRATEGY

ESCALATE_HUMAN

---

Retry Architecture

Maximum automatic retries per stage

2

Example

QA
→ REVISE_COPY
→ Copywriting Agent
→ QA

If the same failure occurs twice

ESCALATE TO HUMAN.

This prevents infinite revision loops.

---

Rework Rules

Do not send work farther upstream than necessary.

Example

Typography problem

QA
→ Visual Prompt Agent

Do NOT

QA
→ Strategy
→ Research

Unsupported marketing claim

QA
→ Strategy

Potentially false source data

QA
→ Research

Use the lowest-cost valid rework path.

---

Human Approval Gate

After QA PASS

QA
→ HUMAN APPROVAL

Human reviewer validates

• Client suitability
• Brand judgment
• Sensitive claims
• Strategic quality
• Commercial readiness

Only after

HUMAN_APPROVED = TRUE

can the Client Delivery Agent execute.

---

Agent 06 — Client Delivery Agent

Responsibility

Package approved assets for delivery.

Can

• Format deliverables
• Organize files
• Create delivery summary
• Prepare client-facing package

Cannot

• Change approved copy
• Change strategy
• Add claims
• Bypass human approval

Start Condition

QA_STATUS = PASS

AND

HUMAN_APPROVED = TRUE

Otherwise

STOP.

---

Workflow State Model

Possible states

RESEARCHING
RESEARCH_VALIDATION
STRATEGY
STRATEGY_VALIDATION
PRODUCTION
QA
REWORK
HUMAN_REVIEW
APPROVED
DELIVERY
COMPLETED
ESCALATED
FAILED

Every workflow instance should have exactly one current state.

---

Context Architecture

Use three context levels.

Global Context

Shared across the entire workflow

• client identity
• campaign objective
• product facts
• brand rules
• prohibited claims

Stage Context

Only information required for the current agent.

Handoff Payload

Structured output from the previous stage.

Avoid forwarding entire conversation histories between every agent.

---

Confidence Architecture

Research Agent

minimum 0.75

Strategy Agent

minimum 0.80

QA

must explicitly PASS.

Confidence alone should not authorize high-risk actions.

---

Failure Conditions

Stop or escalate when

• required source data is missing
• unsupported claims persist
• agent output violates contract
• repeated QA failure exceeds retry limit
• conflicting product information exists
• mandatory human approval is missing
• client requirements contradict internal policy

---

Termination Conditions

SUCCESS

All deliverables approved and delivery completed.

ESCALATED

Human decision required.

FAILED

Workflow cannot continue safely.

Never allow the system to continue indefinitely.

---

Responsibility Matrix

---

Recommended Workflow

Research Agent
↓
Research Validation
↓
Strategy Agent
↓
Strategy Validation
↓
┌────────────────────┐
Copy Agent Visual Agent
└──────────┬─────────┘
↓
Synchronization
↓
QA Agent
↓
PASS / REWORK
↓
Human Approval
↓
Client Delivery
↓
COMPLETED

---

Primary Architecture Insight

MULTI-AGENT SYSTEMS SHOULD NOT BE DESIGNED AS:

“AGENT A TALKS TO AGENT B.”

They should be designed as

ROLE
+
CONTRACT
+
STATE
+
VALIDATION
+
ROUTING
+
FAILURE HANDLING
+
HUMAN CONTROL

The handoff architecture is what turns a collection of agents into a reliable system.