Mahmoud Halat
Six Agent Archetypes

PART IV. THE SIX AGENT ARCHETYPES (DEEP)

Build agents around archetypes; deploy them across domains. The reusability is the point.

Reference implementations to study before building (all v3 additions):

1: Content Generator

What it does: Produces text, image, video, or code outputs against a defined spec.

Across the domains:

  • Domain 3: blog posts, social posts, ad copy, email drafts
  • Domain 4: campaign creative variants, headlines, CTAs
  • Domain 6: outbound email drafts, follow-up sequences
  • Domain 5: AEO-optimized definitional paragraphs

Key design choices:

  • Multi-model routing (Claude for technical, GPT for general, Gemini for multimodal)
  • Brand Governance Agent in the loop (always)
  • Output schema enforcement (so downstream consumers can parse)
  • Iterative refinement (draft → critique → refine; the Reflexion pattern)

Common sub-types:

  • Short-form text generator (headlines, social posts, ad copy)
  • Long-form text generator (blog posts, whitepapers, scripts)
  • Image generator (social cards, illustrations)
  • Video generator (shorts, ad creative)
  • Code generator (landing pages, custom widgets)

2: Knowledge

What it does: Retrieves, synthesizes, and cites from internal and external knowledge sources. Includes RAG (retrieval-augmented generation) over your knowledge base.

Across the domains:

  • Domain 1: synthesizing competitive intelligence from multiple sources
  • Domain 2: pulling positioning research and customer interview transcripts
  • Domain 3: research-first content workflows
  • Domain 6: account research before outreach
  • Domain 7: building synthetic personas from first-party data

Key design choices:

  • Index design (vector DB choice: Pinecone, Weaviate, Qdrant, pgvector)
  • Retrieval strategy (semantic vs. hybrid vs. graph)
  • Citation requirements (every claim should be traceable)
  • Freshness vs. depth tradeoffs
  • Knowledge graph integration (for relationships between entities)

Common sub-types:

  • Internal knowledge synthesizer (Confluence, Notion, Slack archives)
  • Web research agent
  • Account research agent (Clay-style)
  • Customer voice synthesizer (interview transcripts, support tickets)
  • Competitive intelligence agent

3: Localization

What it does: Adapts existing content across languages, regions, segments, regulatory contexts.

Across the domains:

  • Domain 3: translating long-form content for non-English markets
  • Domain 4: regional ad creative adaptation
  • Domain 5: AEO/GEO for non-English LLMs (Yandex, Baidu, regional)
  • Domain 6: regional outbound voice/tone shifts
  • Domain 8: regional benchmarks for measurement

Key design choices:

  • Translation vs. transcreation (translation preserves meaning; transcreation preserves intent)
  • Cultural validation (in-region human review for high-stakes content)
  • Compliance overlay (GDPR, CCPA, regional data rules)
  • Vocabulary localization (terms that work in US ≠ UK ≠ Australia)

Common sub-types:

  • Pure translation agent
  • Transcreation agent (cultural adaptation)
  • Regulatory compliance adapter
  • Persona-segment adapter (enterprise vs. SMB voice)

4: Analyzer

What it does: Examines data and produces insight. The "what's going on?" agent.

Across the domains:

  • Domain 1: anomaly detection in account behavior
  • Domain 4: ad creative performance analysis
  • Domain 6: deal slip prediction, pipeline health scoring
  • Domain 8: attribution analysis, MMM interpretation, incrementality test design and analysis

Key design choices:

  • Statistical rigor (significance testing, confidence intervals)
  • Causal vs. correlational framing (don't claim causation without testing)
  • Anomaly threshold tuning (too sensitive = noise; too loose = missed signals)
  • Insight vs. recommendation distinction (analysis informs decisions; doesn't make them)

Common sub-types:

  • Performance analyzer (campaign / channel / creative)
  • Anomaly detection agent
  • Forecast agent (pipeline, revenue)
  • Attribution interpretation agent
  • Cohort analyzer

5: Planner

What it does: Decomposes a goal into a sequence of actions. The "what should we do?" agent.

Across the domains:

  • Domain 4: campaign planning, channel mix decisions
  • Domain 6: ABM play orchestration, account-level engagement plans
  • Domain 8: experiment design, test prioritization

Key design choices:

  • Decomposition depth (when does the plan stop being useful?)
  • Constraint awareness (budget, timeline, brand boundaries)
  • Re-planning frequency (when do you re-plan vs. execute?)
  • Human approval gates (which decisions need human sign-off?)

Common sub-types:

  • Campaign planner
  • Account-level engagement planner (ABM)
  • Experiment designer
  • Content calendar planner
  • Budget allocation planner

6: Operator

What it does: Executes the plan. Calls APIs, ships content, books meetings, adjusts spend, sends messages.

Across the domains:

  • Domain 3: publishing content to CMS / social platforms
  • Domain 4: launching campaigns, adjusting bids and budgets
  • Domain 6: sending outbound emails, booking meetings, updating CRM
  • Domain 8: pausing underperforming campaigns

Key design choices:

  • Tool permissions (read vs. write; small actions vs. high-stakes)
  • Rate limiting and quotas
  • Failure handling and retry logic
  • Audit logging (every action traced)
  • Reversibility (can this action be undone?)

Common sub-types:

  • Content publisher
  • Ad campaign operator (bid management, budget shifts)
  • Outbound sequence operator (sends, replies, follow-ups)
  • CRM update agent
  • Calendar/meeting operator

Anti-Patterns (when an archetype is the wrong choice)

  • Don't use a Content Generator for strategic positioning. Generators output against a spec, they don't decide what the spec should be. Strategy decisions are Domain 2 / human work.
  • Don't use a Knowledge agent without provenance. A RAG pipeline that returns claims without source attribution is a hallucination factory waiting to fire (cf. Lily Ray's "AI Slop Loop" feedback loop in Domain 5).
  • Don't use a Localization agent for high-stakes regulated content without in-region human review, translation preserves meaning, transcreation preserves intent, but neither catches regulatory risk.
  • Don't use an Analyzer to make causal claims without proper experimental design. Analyzers can spot correlations; causal claims require the Domain 8 incrementality framework.
  • Don't use a Planner without explicit constraints. Open-ended planning produces over-confident multi-step proposals that fail expensively. Always pin budget, timeline, brand boundaries, and approval gates.
  • Don't give an Operator destructive tool access without hardcoded prohibitions. The Replit incident (Domain 0 case studies) is the canonical case study for this failure mode, observability records failures after; enforcement prevents them before.

How to Combine Archetypes

A typical agentic marketing workflow chains 3-5 archetypes:

Chain 1: Account research → personalized outreach (Domain 6 + Domain 1)

"Find an account that's in market, research it, write personalized outreach, send it, follow up."

  1. Analyzer monitors signal feeds (Domain 1) and identifies an in-market account
  2. Knowledge agent does deep research on the account (recent news, leadership, tech stack)
  3. Planner designs the multi-thread engagement strategy (which contacts, what messages, what cadence)
  4. Content Generator drafts the personalized messages
  5. Brand Governance Agent (a Knowledge + Analyzer hybrid) reviews drafts
  6. Operator sends the messages and schedules follow-ups
  7. Analyzer monitors response and triggers re-engagement or human escalation

Reference implementation: Anthropic + Clay (Domain 1 case study). Analyzer + Knowledge layer compresses 60-min research → <5 min per prospect.

Chain 2: Content production pipeline (Domain 3 + Domain 5)

"Take a topic, write a blog post that earns AI citations, ship it, measure citation lift."

  1. Knowledge agent (researcher). Perplexity API + Profound to surface what AI platforms currently cite for the target prompt; identify citation gaps
  2. Analyzer, score competitor content on AEO criteria (answer-block placement, citation density, schema markup); produce a gap report
  3. Planner (outliner), generate the AEO-first content brief (40-60 word answer block, question-led H2s, named-org statistics)
  4. Content Generator (writer). Sonnet 4.5 long-form draft against the brief
  5. Content Generator (editor). Opus 4.7 1M context for high-stakes piece review
  6. Brand Governance Agent (8-agent parallel), voice / grammar / banned terms / legal / persona / sentence structure / terminology / factual grounding; threshold ≥95 = auto-publish
  7. Operator, publish to CMS via API
  8. Analyzer, track citation rate weekly via Profound; feed back to (1) for next round

Reference implementation: langchain-ai/deepagents/examples/content-builder-agent (21.8K stars). The CITABLE-framework B2B SaaS case study (Domain 3), citation rate 8% → 24% in 90 days, uses this exact chain shape.

Chain 3: Measurement loop (Domain 8 + Domain 4)

"Run an incrementality test, decide budget, deploy, measure, iterate."

  1. Planner (experiment designer), define hypothesis, select treatment + control geos, calculate required sample size, set duration
  2. Operator (campaign launcher), deploy via Meta/Google Ads API; suppress treatment in control geos
  3. Knowledge (data sync), pull spend, exposure, conversions from ad platforms + CRM into the warehouse
  4. Analyzer (causal interpreter), run synthetic-control or geo-test analysis; produce incremental ROAS per channel
  5. Planner (budget reallocator), feed incrementality output as Bayesian priors into MMM (Recast / Lifesight / Robyn); recommend budget reallocation
  6. Operator (budget adjuster), push new budget allocations to ad platforms
  7. Analyzer (anomaly watcher), alert on CAC drift > tolerance; loop back to (1) when next channel needs validation

Reference implementation: Soft Surroundings (Domain 8 case), cut retargeting 52% based on incrementality test → +17% revenue MoM, +12% YoY. PODS (Recast), moved from 2x/year MMM to weekly updates → +181% Google Non-Brand spend after go-dark test.

Chain 4: Brand governance (Domain 0 + Domain 2 + Domain 3)

"Every public-facing AI output gets reviewed before publish."

This chain runs as a service consumed by Chains 1, 2, and others, not a standalone workflow.

  1. Orchestrator (a thin Knowledge agent), receives draft + metadata (target persona, channel, brand spec version)
  2. 8 specialized Analyzer agents in parallel, each with a single audit dimension and its own context window:
    • Voice auditor (matches brand voice axes)
    • Grammar auditor
    • Banned-terms auditor (against the ontological don't-list)
    • Legal-claims auditor (against approved-claims registry)
    • Persona-fit auditor (target persona consumption test)
    • Sentence-structure auditor (banned patterns: "In today's fast-paced world...")
    • Terminology auditor (terms-we-own enforcement)
    • Factual-grounding auditor (citation verification)
  3. Aggregator (Analyzer), produces single Brand Governance Score 0-100
  4. Decision gate (Operator), score ≥95 auto-publish; 85-94 human review queue; <85 block + return to source agent for rework
  5. Audit logger (Operator), every decision logged with {brief, draft, governance_score, flagged_violations, approver, brand_spec_version} for AgentOps observability + regulatory audit trail

Reference implementation: Brand.ai (Lyft, Opendoor), enterprise brand-compliance costs $5M/yr → small fraction; one copywriter manages 600 pieces of content (Domain 2 case). The Animalz "Claude Code style-guide" pattern published in 2025 + anthropics/skills brand-guidelines repo.

What every chain shares

  • Knowledge agents inject context at the start of the chain
  • Analyzer agents validate quality at multiple checkpoints, never just at the end
  • Planner agents produce explicit, constrained plans, not open-ended "figure it out"
  • Brand Governance Agent gates every public-facing output (Chains 1 + 2 cite to Chain 4 as a sub-routine)
  • Operator agents are sandboxed: they can only call tools that have been pre-approved, with hardcoded prohibitions on destructive actions (the Replit lesson)
  • Audit logs are immutable and feed AgentOps observability for drift detection (Domain 0)

Each archetype is reusable across these chains. Build once, deploy many.

Future Expansion (v4 target)

This file is the v3 baseline (current). See research-plan.md for methodology and master plan. Per-file checklist:

  • Primary sources to mine, highest-signal reports, papers, podcasts, expert posts to extract from (with URLs and dates).
  • Claims to validate / refresh, statistics that need primary-source verification or are stale.
  • Case studies to add, real deployments with named clients, numbers, before/after data.
  • Tactical playbooks to add, build-this-yourself walkthroughs, prompts, configs, decision trees.
  • Tooling deep-dives, head-to-head comparisons, pricing notes, integration gotchas, switching costs.
  • Open questions, unresolved debates worth tracking.
  • Cross-references, links to Mahmoud's other skills, related domain files, external knowledge bases.
Continue reading
Resource Library