← Field Guide  ·  Tier 1 · Foundational Industry & Research

Academic Foundations

Key papers behind every major pattern

Foundational · arXiv / conferences17 patterns
In Plain English

This page shows every pattern that Academic Foundations has published or written about: 17 in total. Each one includes a simple picture of how it works, a short explanation, an example of when to use it, and its trade-offs (the good parts and the not-so-good parts).

Reading patterns from one source at a time matters because it lets you see how a single company or team thinks about building AI agents, and compare their ideas side by side. If a pattern here also appears in the bigger Deep Catalogue, you will see a link so you can read the full, detailed version.

An open reference book with a pencil in the fold under warm lamp light

1D.01ReAct

ReAct (Yao et al. ICLR 2023) Thought Action Observation → Thought

Interleave reasoning traces with tool actions. Yao et al. ICLR 2023.

Use caseDefault reasoning pattern for most tool-using agents.
Trade-offsTransparent · Higher latency than single-shot.

See 1A.08, 1B.10

1D.02Tree-of-Thoughts

TREE-OF-THOUGHTS Problem Thought A score 0.7 Thought B prune Thought C score 0.8 ✓

Generate multiple candidate thoughts per step; evaluate; prune; continue best. Yao et al. NeurIPS 2023.

Use caseCombinatorial reasoning, design generation, puzzles.
Trade-offsQuality lift on branching problems · Multiplicative cost.

1D.03Reflexion

REFLEXION Agent Try Feedback Reflect Retryattemptevaluatecritiqueretry

Agent reflects on failed trajectories, stores verbal self-critique in memory, uses it in subsequent attempts. Shinn et al. NeurIPS 2023.

Use caseTasks with feedback signal; coding, decision-making.
Trade-offsIterative improvement · Requires failure signal.

1D.04Self-Refine

SELF-REFINE Same Model Generate Self-Critique Refinegencritiquerefine

Single model iteratively generates, self-critiques, and revises. Madaan et al. NeurIPS 2023.

Use caseWriting, code, dialogue response — anywhere revision improves quality.
Trade-offsMeasurable quality gains · Doubled-or-more cost.

1D.05Plan-and-Solve

PLAN-AND-SOLVE Read Task Devise Plan Execute Plan Answer

Explicit "devise a plan, then carry out the plan" instruction that outperforms zero-shot CoT. Wang et al. 2023.

Use caseArithmetic, commonsense, symbolic reasoning.
Trade-offsBetter than raw CoT · Still single-agent in scope.

1D.06ReWOO

ReWOO Planner Plan Workers ∥ Solver

Reasoning WithOut Observation: full plan before any tool calls; workers execute; solver composes. Xu et al. 2023.

Use caseLow-adaptation workflows where re-planning is expensive.
Trade-offsCheaper than ReAct on predictable tasks · Less adaptive.

See 1C.03

1D.07Self-Consistency

SELF-CONSISTENCY Promptsame task N× Chain 1 Chain 2 Chain 3Take majority answer.

Sample multiple reasoning chains; take majority answer. Wang et al. ICLR 2023.

Use caseArithmetic, reasoning benchmarks where answers are discrete.
Trade-offsAccuracy lift without fine-tuning · N× token cost.

See 1A.05 (voting)

1D.08LATS (Language Agent Tree Search)

LATS Root Branch A UCB score Branch B expand Branch C best

Monte-Carlo Tree Search over reasoning + acting + reflection. Zhou et al. 2023.

Use caseComplex decision problems where exploring alternatives pays off.
Trade-offsState-of-the-art on some benchmarks · Expensive.

1D.09Toolformer

TOOLFORMER Seed Examples Self-Annotate Filter by Gain Tool-Augmented Model

Self-taught tool use: model learns when and how to call APIs from a small set of demonstrations. Schick et al. NeurIPS 2023.

Use caseExtending model capability with external tools without fine-tuning on tool-use data.
Trade-offsTool-calling without heavy supervision · Upfront self-annotation needed.

1D.10Gorilla

GORILLA Query Tool Retriever Top-K Tools Model Call

LLM trained to invoke massive API catalogs via retrieval-augmented tool calling. Patil et al. 2023.

Use caseEnterprise agents with hundreds of tools.
Trade-offsScales tool count · Requires tool-retrieval pipeline.

Cross-ref: Deep Catalogue TOO-DEC-02 →

1D.11AutoGen

AUTOGEN Chat Manager User Proxy Assistant Critic

Framework for multi-agent conversation between configurable agents. Wu et al. 2023.

Use casePrototyping multi-agent systems with flexible roles.
Trade-offsFast to wire up · Conversation cost can spiral.

1D.12MetaGPT

METAGPT Requirements PM specs Architect design Engineer code

Role-based multi-agent system simulating a software development team. Hong et al. 2023.

Use caseStructured software tasks with clear role division.
Trade-offsRich output (code + docs + tests) · Rigid role structure.

1D.13Voyager

VOYAGER Agentskill library Explore Skill Add to Library Reusetrysucceedstorecall

Lifelong learning agent in Minecraft with a growing skill library. Wang et al. 2023.

Use caseOpen-ended environments where skills accumulate.
Trade-offsReusable skill library · Narrow demonstration domain.

1D.14Generative Agents

GENERATIVE AGENTS Memory Streamraw observations Reflectionsynthesise insights Planningday-by-day

Simulation agents with memory streams, reflection, and planning: believable human behaviour. Park et al. 2023.

Use caseSimulation, NPCs, social experiments, training environments.
Trade-offsEmergent believable behaviour · Expensive for production.

1D.15Self-RAG

SELF-RAG Query Retrieve? Generate + Critique Answer

Model learns to decide when to retrieve, critique retrievals, and critique its own output. Asai et al. ICLR 2024.

Use caseFactual Q&A where over-retrieval hurts fluency and under-retrieval hurts accuracy.
Trade-offsAdaptive retrieval · Training-intensive.

Cross-ref: Deep Catalogue RAG-ACT-02 →

1D.16CRAG (Corrective RAG)

CORRECTIVE RAG Retrieve Evaluate Relevance Web Search if Low Generate

Evaluates retrieval quality; triggers web search or rewrite when retrieved context is insufficient. Yan et al. 2024.

Use caseProduction RAG where corpus coverage is uneven.
Trade-offsResilient to retrieval gaps · Additional evaluation step.

1D.17Graph-of-Thoughts

GRAPH-OF-THOUGHTS Input Node merge A+B Node refine Node back-edge

Reasoning state as an arbitrary DAG: supports aggregation, refinement, back-tracking. Besta et al. 2024.

Use caseProblems where combining intermediate results matters.
Trade-offsMore expressive than ToT · Scheduler complexity.
Also in Tier 1
Anthropic 8Google Cloud 12Antonio Gulli 17