# Multi-Agent Handoff Architecture ## Executive Summary Recommended topology: Research → Research Validation → Strategy → Strategy Approval → Copy + Visual Prompt → QA → Human Approval → Client Delivery The proposed six-agent system is viable, but direct sequential handoffs would create three major risks: 1. Weak research propagating into strategy. 2. Copy and visual agents receiving different interpretations. 3. QA repeatedly rewriting work without convergence. Production Status: READY WITH CONDITIONS --- ## Agent Registry ### A1 — Research Agent Responsibility: Produce evidence-backed campaign research. May: Collect and structure verified information. May Not: Create final strategy. Invent unsupported customer facts. Output: RESEARCH_BRIEF --- ### A2 — Strategy Agent Responsibility: Convert approved research into campaign strategy. May Not: Replace missing research with assumptions presented as facts. Input: APPROVED RESEARCH_BRIEF Output: STRATEGY_BRIEF --- ### A3 — Copywriting Agent Responsibility: Create campaign copy from approved strategy. May Not: Change target audience, offer, campaign promise, or factual claims. --- ### A4 — Visual Prompt Agent Responsibility: Translate approved creative direction into image-generation instructions. May Not: Create a different campaign strategy. --- ### A5 — QA Agent Responsibility: Validate contract compliance. May: Reject specific defects. May Not: Rewrite valid work or redesign approved strategy. --- ### A6 — Client Delivery Agent Responsibility: Package approved artifacts. May Not: Modify substantive content. May Not: Deliver without final approval. --- ## Primary Artifact Chain RESEARCH_BRIEF → STRATEGY_BRIEF → COPY_PACKAGE + VISUAL_PACKAGE → QA_REPORT → APPROVAL_RECORD → CLIENT_PACKAGE --- ## Research → Strategy Handoff Required Fields: • Business objective • Target audience evidence • Competitor observations • Market findings • Source references • Known unknowns • Unsupported assumptions Acceptance Criteria: • Required sections complete • Important claims sourced • Contradictory evidence identified • Unverified assumptions clearly labeled Hard Gate: HG-2 SOURCE GROUNDING FAILURE If evidence supporting a major strategic claim is absent: HANDOFF BLOCKED. Strategy Agent must not compensate by inventing facts. --- ## Strategy → Copy Handoff Context Package: MUST KNOW • Campaign objective • Target audience • Core proposition • Messaging hierarchy • Approved claims • CTA • Restrictions MAY KNOW • Selected supporting research DO NOT NEED • Full raw research conversation • Rejected strategic alternatives This reduces context pollution. --- ## Strategy → Visual Handoff Visual Agent receives the same approved strategy version as Copywriting. Artifact Reference: STRATEGY_BRIEF v1.0 APPROVED This prevents copy and visual production from diverging. --- ## Parallel Execution After strategy approval: Copywriting Agent and Visual Prompt Agent may operate in parallel. Neither depends on the other's draft. Both depend on the same approved strategy. --- ## QA Gate QA checks: Copy: • Strategy alignment • Unsupported claims • CTA compliance • Required deliverables Visual: • Strategy alignment • Product consistency • Required format • Brand restrictions QA does not perform subjective rewrites without a contract violation. --- ## Defect Rules S1 Minor: Targeted repair only. S2 Material: Return affected artifact only. S3 Critical: Block workflow. Example: Copy has unsupported statistics. Return Copy only. Do not rerun: Research Strategy Visual unless the defect originates upstream. --- ## QA Loop Control Maximum QA revision cycles: 2 Cycle 1: Targeted repair. Cycle 2: Final repair. If the same S2/S3 defect remains after cycle 2: ESCALATE TO HUMAN. No third automatic QA loop. --- ## Human Approval Required before: CLIENT_DELIVERY Human receives: • Strategy summary • Final copy • Visual prompts • QA result • Outstanding assumptions Approval options: APPROVE REQUEST TARGETED REVISION REJECT --- ## State Machine INTAKE → RESEARCH → RESEARCH_VALIDATION → STRATEGY → STRATEGY_APPROVAL → PRODUCTION → QA → HUMAN_APPROVAL → DELIVERY → COMPLETED Failure paths: RESEARCH_VALIDATION → RESEARCH_REVISION QA → TARGETED_REVISION Retry exhausted → ESCALATED Critical unresolved defect → FAILED_SAFE --- ## Critical Architecture Rule Client Delivery Agent receives only APPROVED artifacts. It cannot convert DRAFT or REVISION_REQUIRED artifacts into FINAL status. --- ## Production Recommendation READY WITH CONDITIONS Required controls: 1. Add evidence gate before Strategy. 2. Use one approved Strategy artifact as source of truth. 3. Limit QA to two cycles. 4. Restrict QA from rewriting valid work. 5. Require human approval before delivery. 6. Version all strategy-dependent artifacts.