← Field Guide  ·  Tier 1 · Foundational Industry & Research

Antonio Gulli

Agentic Design Patterns (Google/Springer, Dec 2025) — 21 chapters

Foundational · book17 patterns
In Plain English

This page shows every pattern that Antonio Gulli 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

1C.01Reflection

REFLECTION Agentgen ⇄ critic Generate Critique Revise Exitgencritiquerevisedone

Agent critiques its own output against explicit criteria and revises before returning. Gulli Ch. 4.

Use caseHigh-stakes outputs, code generation, regulated-content drafting.
Trade-offsMeasurable quality lift · Doubled-or-more token cost.

Cross-ref: Deep Catalogue REA-RFL-01; overlaps 1A.07, 1B.05 →

1C.02Tool Use

TOOL USE Tool Schema Agent Validator Execute Typed Return

Every tool is a typed function with declared input/output schemas. Invocations are validated; results are structured. Gulli Ch. 5.

Use caseAny agent calling APIs, databases, file systems, other services.
Trade-offsReliable invocation, static analysis possible · Schema authorship cost.

Cross-ref: Deep Catalogue TOO-DEC-01 →

1C.03Planning (Plan-and-Execute / ReWOO)

PLAN-AND-EXECUTE Planner DAG Worker ∥ Solver Answer

Generate a complete plan up front; execute steps (often in parallel); re-plan only on real break. Gulli Ch. 6. Originates Xu et al. ReWOO 2023.

Use casePredictable-structure tasks, research pipelines, low-novelty automation.
Trade-offsCheaper per-task · Less adaptive than ReAct.

Cross-ref: Deep Catalogue REA-PLN-02 →

1C.04Multi-Agent Collaboration

COLLABORATION Protocolturn-taking Role A Role B Role C

Specialists with distinct roles work together, often with explicit protocols for turn-taking, handoff, and conflict resolution. Gulli Ch. 7.

Use caseRole-based teams (PM + engineer + designer), simulation, debate.
Trade-offsRicher problem-solving · Coordination overhead.

1C.05Memory Management

MEMORY STRATA Short-Termworking · scoped to task Episodicwhat happened Semanticwhat is true Proceduralhow to do things

Stratified memory: short-term (working), long-term episodic, semantic, procedural. Each stratum has retention, retrieval, access policy. Gulli Ch. 8.

Use caseLong-running agents, personalisation, recurring user interactions.
Trade-offsClean retrieval semantics · Architectural complexity.

Cross-ref: Deep Catalogue MEM-PER-01, MEM-RFL-01 →

1C.06Learning & Adaptation

LEARNING MODES In-Contextfew-shot at inference Parametricfine-tune weights FeedbackRLHF · reflection

In-context learning (few-shot), parametric (fine-tuning), feedback-driven (RLHF / reflection traces). Gulli Ch. 11.

Use casePersonalisation, domain adaptation, continuous improvement.
Trade-offsBetter fit over time · Drift risk if feedback signal poor.

1C.07Model Context Protocol (MCP)

MCP ENVELOPE Agent MCP Client MCP Server Tool / Data

Standard protocol for agents to connect with tools, data sources, and other agents via a uniform envelope. Gulli Ch. 10.

Use caseEnterprise integrations, cross-vendor ecosystems, tool marketplaces.
Trade-offsPortability · Standards are young; some features not yet covered.

Cross-ref: Deep Catalogue A2A-DEC-01

1C.08Goal Setting & Monitoring

GOAL MONITORING Declare Goal Sub-goals Progress Done Criteria Close

Explicit goal declaration, sub-goal tracking, progress measurement against declared criteria. Gulli Ch. 11.

Use caseLong-running agents, multi-step missions, autonomous operations.
Trade-offsVisibility into progress · Measurement overhead.

Cross-ref: Deep Catalogue GOA-DSN-01, GOA-RFL-01 →

1C.09Exception Handling & Recovery

EXCEPTION HANDLING Tool Wrapper Attempt Retry+Backoff Circuit Break Degradetryretrytripfallback

Bounded retries, backoff, circuit breaking, graceful degradation paths. Gulli Ch. 12.

Use caseProduction agents with external dependencies, high-availability requirements.
Trade-offsResilience · Library/framework dependencies.

Cross-ref: Deep Catalogue EXC-ACT-01, EXC-OBS-01 →

1C.10Knowledge Retrieval (RAG)

RAG Query Retrieve Rerank Generate Cited Answer

Retrieve relevant passages from authoritative indexes; ground generation with citations. Gulli Ch. 14.

Use caseDomain-expert agents, support over product knowledge, regulatory/legal agents.
Trade-offsReduced hallucination · Retrieval pipeline to maintain.

Cross-ref: Deep Catalogue RAG-PER-01, RAG-ACT-01, RAG-ACT-02 →

1C.11Inter-Agent Communication (A2A)

AGENT-TO-AGENT Agent X A2A Envelopecapabilities Agent Y Result

Standard envelopes for agent↔agent messaging with capability handshake and versioning. Gulli Ch. 15.

Use caseMulti-agent platforms, cross-vendor agent integration.
Trade-offsInteroperability · Standards still evolving.

1C.12Resource-Aware Optimisation

BUDGET CONTROL Task Startbudgetdeclared Step Budget Check Continue Halt

Track and cap per-task budgets: tokens, cost, latency, tool calls. Gulli Ch. 16.

Use caseHigh-volume agents, multi-tenant platforms, cost-sensitive workloads.
Trade-offsBounded worst-case cost · May cut off legitimate long work.

Cross-ref: Deep Catalogue EXC-ACT-01, PLN-DEC-01 →

1C.13Reasoning Techniques

REASONING PATTERNS Problem Chain (CoT) linear Tree (ToT) branching Graph (GoT) DAG

Chain-of-Thought (linear), Tree-of-Thoughts (branching), Graph-of-Thoughts (arbitrary DAG), LATS (tree search). Gulli Ch. 17.

Use caseMathematical reasoning, design problems, combinatorial search.
Trade-offsBetter accuracy on branching problems · Multiplicative token cost.

Cross-ref: Deep Catalogue PLN-PLN-02 →

1C.14Guardrails & Safety

GUARDRAIL STACK Input Filterinjection scan Context Separationdata ≠ instructions Output FilterPII · policy Kill-Switchemergency stop

Input filters, structural separation of data from instructions, output filters, topic confinement, kill-switch. Gulli Ch. 18.

Use caseAny production agent; mandatory for regulated sectors.
Trade-offsDefence in depth · Multiple layers to maintain.

Cross-ref: Deep Catalogue GRD-PER-01, GRD-ACT-01, GRD-ACT-02, GRD-END-01 →

1C.15Evaluation & Monitoring

EVAL LOOP Golden Set Red Team Live Telemetry Spot-Check Feedback

Golden-set evals, red-teaming, human spot-check, live telemetry with anomaly detection. Gulli Ch. 19.

Use caseAll production agents; mandatory under IMDA MGF and EU AI Act.
Trade-offsRegression protection · Test authorship and operational cost.

Cross-ref: Deep Catalogue EVA-DSN-01, EVA-EVL-01, EVA-ITR-01, EVA-OPR-01 →

1C.16Prioritisation

PRIORITISATION Queue Scoreimpact×urgency×cost Rank Execute Top

Score candidate tasks by impact × urgency × cost; execute in ranked order; re-score on new info. Gulli Ch. 20.

Use caseAgents juggling concurrent tasks — inbox, monitoring, research.
Trade-offsResource efficiency · Scoring function tuning.

Cross-ref: Deep Catalogue PLN-DEC-01 →

1C.17Exploration & Discovery

EXPLORATION Explorer Hypothesise Test Observe Updateproposetestlearniterate

Deliberate search of solution space, hypothesis generation, novelty-seeking behaviour. Gulli Ch. 21.

Use caseResearch agents, scientific exploration, creative problem-solving.
Trade-offsNovel solutions · Hard to bound and evaluate.
Also in Tier 1
Anthropic 8Google Cloud 12Academic Foundations 17