Skip to content

Agentic Pattern Vocabulary Crosswalk

Crosswalk that maps four agentic pattern vocabularies — Ng's four, Anthropic's five, Gulli's twenty-one, CoALA's three axes — to this site's canonical pages.

External readers and AI agents arrive with whichever vocabulary their source used — "Reflection" from Andrew Ng, "Orchestrator-Workers" from Anthropic, "Goal Setting" from Antonio Gulli, "structured action space" from CoALA. This crosswalk is a navigation aid that resolves each external term to the site's canonical page for the same concept. It is not a second authoritative definition of any pattern: every row leads with the link, and the body lives on the linked page.

For the cross-pattern trade-off view (token cost, latency, blast radius, verification cost), use the sibling Pattern Selection Map. For the broader cross-vendor terminology discussion (workflow vs agent vs assistant vs RAG pipeline), see Agent Terminology Disambiguation. For an Anthropic-only mapping with augmented-LLM context, see Anthropic's Effective Agents Framework.

How to Read the Tables

Each row gives the external term, a one-line scope-difference note (because the mappings are rarely 1:1), and the site's canonical page. A blank "Scope difference" cell means the external term and the site page name the same concept at the same scope. Coverage as of last_reviewed.

Andrew Ng's Four Agentic Design Patterns

Andrew Ng's 2024 DeepLearning.AI series names four patterns that improve agentic-coding results across model families, drawing from a corpus of papers including AutoGen, ReAct, and ChatDev (DeepLearning.AI — Agentic Design Patterns Part 1; covered on-site at Anthropic's Effective Agents Framework and Agent Terminology Disambiguation).

Ng's pattern Scope difference This site's canonical page
Reflection Ng frames reflection as a self-evaluation step inside an agent loop. The site splits this across a code-review-loop variant and a generator/critic split. Agent Self-Review Loop, Evaluator-Optimizer
Tool Use Ng treats tool use as a single pattern; the site decomposes it into tool design, advanced tool use, MCP, and skills. Advanced Tool Use, Tool Engineering
Planning Ng's planning is a runtime "decompose first, then execute" loop; the site's plan-first-loop is the workflow shape, while the cognitive split is a separate pattern. Plan-First Loop, Cognitive Reasoning vs Execution Separation
Multi-Agent Collaboration Ng treats multi-agent as one pattern; the site indexes it as a whole section organised by topology. Multi-Agent (section index), Multi-Agent Topology Taxonomy

Anthropic's Five Workflow Patterns

Anthropic's Building Effective Agents names five workflow patterns — control-flow defined in code — plus an autonomous-agent loop where the model owns control flow (Anthropic — Building Effective Agents). The full taxonomy mapping with augmented-LLM context lives at Anthropic's Effective Agents Framework; the rows below are the navigation lookup.

Anthropic pattern Scope difference This site's canonical page
Prompt chaining Prompt Chaining
Routing Anthropic names the classify-then-dispatch shape. The site's canonical page emphasises the cost-aware "cheapest model that solves it" variant. Parsimonious Agent Routing
Parallelization Anthropic covers sectioning (independent subtasks) and voting (consensus). The site's fan-out-synthesis is the sectioning variant; voting is a separate pattern. Fan-Out Synthesis, Voting Ensemble Pattern
Orchestrator-workers Orchestrator-Worker
Evaluator-optimizer Evaluator-Optimizer

For the autonomous-agent layer above these five workflows, see Goal-Driven Autonomous Loop, Loop Strategy Spectrum, and Agentless vs Autonomous.

Antonio Gulli's Twenty-One Agentic Design Patterns

Antonio Gulli's Agentic Design Patterns book organises 21 patterns across Core (8), Advanced (6), and Specialized (7) categories. A reference implementation in LangChain + Gemini covers all 21 (josephsenior/Agentic-Design-Patterns). Coverage flags below: covered = at least one direct canonical page, partial = adjacent pages but no first-class treatment, gap = no direct coverage as of last_reviewed.

Gulli's pattern Scope difference This site's canonical page Status
Prompt Chaining Prompt Chaining covered
Routing Same scope difference as Anthropic's row above. Parsimonious Agent Routing covered
Parallelization Same scope difference as Anthropic's row above. Fan-Out Synthesis covered
Reflection Same scope difference as Ng's row above. Agent Self-Review Loop covered
Tool Use Advanced Tool Use covered
Planning Plan-First Loop covered
Multi-Agent Multi-Agent (section index) covered
Memory Management Gulli treats memory as one pattern; the site decomposes it into scope/temporal patterns and a CoALA-aligned taxonomy. Agent Memory Patterns, CoALA Memory Taxonomy Classifier covered
Learning and Adaptation Gulli covers cross-task adaptation; the site's page focuses on layered continual learning across model and harness. Continual Learning Layers covered
Model Context Protocol (MCP) MCP Protocol covered
Goal Setting and Monitoring Goal Monitoring and Progress Tracking covered
Exception Handling and Recovery Exception Handling and Recovery Patterns covered
Human-in-the-Loop Gulli covers human-in-the-loop broadly; the site has a workflow page and a security-side confirmation-gates page. Human-in-the-Loop, Human-in-the-Loop Confirmation Gates covered
Knowledge Retrieval (RAG) Gulli treats RAG as a single pattern; the site's canonical RAG page frames it as on-demand / JIT context retrieval (and declares "RAG" an explicit alias). Retrieval-Augmented Agent Workflows covered
Inter-Agent Communication (A2A) A2A Protocol covered
Resource-Aware Optimization Gulli covers cost/latency optimisation generally; the site's coverage is the cost-aware-design page plus the routing page above. Cost-Aware Agent Design covered
Reasoning Techniques Gulli covers chain-of-thought, ReAct, and tree-of-thoughts as reasoning patterns. The site treats reasoning as a cognitive-execution split plus a three-spaces decomposition. Cognitive Reasoning vs Execution Separation, Three Reasoning Spaces covered
Guardrails / Safety Patterns Gulli covers safety as one pattern; the site has a four-layer taxonomy and a defence-in-depth page that decompose it further. Four-Layer Agent Security Taxonomy, Defense-in-Depth Agent Safety covered
Evaluation and Monitoring Gulli covers evaluation as one pattern; the site separates evaluator-as-pattern from eval-as-workflow. Evaluator-Optimizer, LLM-as-Judge Evaluation covered
Prioritization Gulli treats prioritization as next-action ranking; the site's page focuses on rank-vs-routing-vs-scheduling and dependency-aware signals. In-Agent Task Prioritization covered
Exploration and Discovery Gulli frames exploration as a search-and-discover pattern; the site's coverage is the best-of-n delegation page and a discovery-only refactor workflow. Recursive Best-of-N Delegation, Discovery-Only Refactor Pass covered

CoALA's Three Axes

CoALA (Cognitive Architectures for Language Agents) is a peer-reviewed conceptual framework that organises language agents along three axes (arxiv:2309.02427; on-site at Cognitive Architectures for Language Agents (CoALA)). Each axis has its own canonical page.

CoALA axis Scope difference This site's canonical page
Memory (four types: working, episodic, semantic, procedural) The CoALA classifier page does the missing-slot diagnostic; the broader memory-patterns page covers the scope/temporal choices. CoALA Memory Taxonomy Classifier, Agent Memory Patterns
Action space (internal vs external) CoALA Structured Action Space
Decision-making loop (propose → evaluate → select → act) The CoALA page frames the loop as an orchestration vocabulary, not a runtime prescription. CoALA Decision-Making Loop

Why It Works

Practitioners arrive at agent-pattern material with the vocabulary their training, vendor docs, or community catalogue used — and that vocabulary rarely matches the mechanism-organised vocabulary a build-out composes. The mechanism is cognitive offloading of the cross-vocabulary lookup step, the same one the sibling OWASP LLM Top 10 (2025) Agent Security Crosswalk documents: "Practitioners arrive at security material with the vocabulary their training, compliance reviews, and tooling use… A crosswalk closes the gap by giving readers a stable mapping from the framework names they searched to the mechanism-organised pages…" Without the crosswalk, a reader arriving with one vocabulary (Ng's "Reflection", say) cannot find the site's coverage because the site indexes patterns by mechanism (agent-self-review-loop, evaluator-optimizer) and by topology (multi-agent/orchestrator-worker), not by external taxonomy names. The crosswalk also captures the GEO disambiguation-hub surface and avoids duplicate idea issues being filed for patterns the site already covers under different names.

The shape is deliberate: leading every row with the scope-difference note and the link — not with a definition in the site's own words — protects the canonicality contract (.claude/rules/canonicality.md) that reference trees own each concept. The page summarises and links; it does not restate.

When This Backfires

A crosswalk is a discovery aid, not a recommendation engine, and four failure modes make it actively misleading if treated as one.

  • 1:1 mapping fallacy. Reading the tables as drop-in equivalences ("Ng's Reflection = Anthropic's Evaluator-Optimizer = Gulli's Reflection") loses the scope differences that the per-row notes flag. A 2025 paper argued "agent" is "diluted beyond utility" and proposed multidimensional characterisation rather than single definitions (arxiv:2508.05338, cited on-site at Agent Terminology Disambiguation); the same dilution affects pattern names. Use the rows for navigation, then read each linked canonical page for the actual mechanism and trade-offs.
  • Coverage flag staleness. Pattern pages get renamed, deprecated, or split over time. The "gap" / "covered" status flags are dated by last_reviewed; a flag that read "covered" in 2026-06 may be wrong six months later. The full-audit pipeline refreshes the date on every per-page audit, but a reader landing here years later should treat coverage as a snapshot, not a current index.
  • Vocabulary churn outpaces refresh. Gulli's 21-pattern catalogue is a 2025 snapshot. A future edition (or a competing 21-pattern catalogue) may rename or re-split patterns. Anthropic's framework expanded from a December 2024 blog post to a 2025 eBook with case studies and skills coverage (eBook landing). The crosswalk anchors to a specific snapshot per last_reviewed; treat it as that snapshot.
  • Restatement creep. If a future edit adds a paragraph defining "Reflection" or "Routing" on this page in the site's own words, the crosswalk has crossed the line from navigation to a second authoritative definition — the canonicality defect #7736 names. Keep every row as a link plus a one-line scope difference; the body lives on the linked page.

Key Takeaways

  • A vocabulary crosswalk maps external taxonomy names to canonical site pages so readers find coverage by whichever name they arrived with; it is navigation, not a competing taxonomy.
  • Ng's four (Reflection, Tool Use, Planning, Multi-Agent), Anthropic's five (chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer), Gulli's 21 (Core/Advanced/Specialized), and CoALA's three axes (memory, action, decision-loop) all map cleanly onto this site's mechanism-organised pages with documented scope differences.
  • All 21 of Gulli's patterns are covered on the site as of last_reviewed, with scope differences noted per row (e.g., Memory decomposed across memory-patterns and CoALA-classifier; RAG mapped to the canonical Retrieval-Augmented Agent Workflows reference).
  • Every row leads with the canonical page link plus a one-line scope-difference note — there is no definition of any pattern in this page's own words, by design.
  • For pattern trade-offs (token cost, latency, blast radius), pair this crosswalk with the sibling Pattern Selection Map; for cross-vendor terminology (workflow vs agent vs assistant), pair it with Agent Terminology Disambiguation.
Established — multiple independent sources Last reviewed: 2026-06-16
Feedback