The most fundamental building block
These agents perceive their environment, evaluate actions against goals and constraints, choose, and execute — without continuous supervision — learning by updating state for the next loop. They realize Chapter 1's cognitive loop: perception → cognition → action → learning.
Key concept · From reactive to autonomous
The leap is enhanced perception — beyond the raw message, the agent adds situational awareness (system load, time of day, user tier, active alerts, human-agent availability). Its cognition then scores candidate strategies (full-autonomous resolution, immediate escalation, guided resolution) across weighted axes — autonomy level, urgency, complexity, escalation threshold — and picks the highest. That's what separates it from a simple reactive script.
Orchestrating dynamic, multi-step workflows
Where a decision-maker answers a single query, a Planning agent orchestrates a whole product launch. The core technique is hierarchical decomposition — turning an abstract goal into a tree of smaller, concrete sub-goals. Two complementary paradigms:
Symbolic planning
Formal representations — STRIPS and PDDL — define states, actions, goals in structured languages; algorithms search for optimal sequences. Precise with optimality guarantees, but needs comprehensive pre-modeling; brittle in novel situations.
LLM-powered dynamic planning
Techniques like Tree-of-Thought and Self-Ask generate and explore solution paths in natural language — no rigid pre-programming, highly adaptable to novel/emergent requirements. Leans on chain-of-thought prompting from Chapter 3.
Implementation rests on three capabilities: strategic task decomposition (sequential + parallel subtasks, dependencies), dynamic execution & monitoring (track milestones, trigger revisions), and adaptive intelligence (revise plans as conditions change, drawing on memory of past projects).
| Capability | Decision-maker | Planning agent |
|---|---|---|
| Focus | Immediate, tactical | Strategic, long-term |
| Scope | Single decisions | Multi-step, coordinated |
| Adaptation | Heuristics / conditional logic | Dynamic revision & learning loops |
| Memory use | Within current session | Persistent, cross-project |
Continuity across sessions
Many agents live only in the present. Memory-Augmented agents preserve context over time using three memory systems inspired by human cognition:
Key concept · Three memory types
- Working memory — short-term, in-session; held directly in the LLM prompt (recent turns, current goals). Transient — cleared at session end or token limit.
- Episodic memory — timestamped history of interactions/events; recalled via vector similarity search. Enables continuity & personalization.
- Semantic memory — structured factual knowledge (specs, regulations, facts) from documents/APIs/DBs; queried by keyword or semantic search. Grounds claims in authoritative data.
Skipping the right memory has distinct failure modes: no working memory → loss of conversational coherence; no episodic → each session treated as new (broken personalization); no semantic → hallucinated or stale answers. The chapter's healthcare-assistant case study shows all three working together for personalized, grounded support.
Complementary, not mutually exclusive
- Decision-making
- Real-time responsiveness (chatbots, alerting, triage). Trade-off: speed can raise hallucination/unsafe-action risk; lacks strategic depth.
- Planning
- Complex multi-step goals (orchestration, workflow automation). Trade-off: slower; needs robust monitoring & recovery.
- Memory-augmented
- Continuity & personalization (CRM, assistants). Trade-off: added storage/retrieval complexity and governance.
Their true potential emerges through integration — combined, they realize the full cognitive loop and move agents from reactive behavior toward long-term, context-aware intelligence.
Chapter 5 quiz
Fifteen questions on the three cognitive architectures. Answer first, then expand Show answer.
Part A · Multiple choice
Which memory type is held directly in the LLM prompt and cleared at session end or the token limit?
- A Episodic memory
- B Semantic memory
- C Working memory
- D Procedural memory
Show answer
C — working memory. Short-term, in-session. Episodic is timestamped history; semantic is structured factual knowledge.
STRIPS and PDDL are associated with which planning paradigm?
- A LLM-powered dynamic planning
- B Symbolic planning
- C Tree-of-Thought planning
- D Reactive planning
Show answer
B — symbolic planning. Formal languages for states/actions/goals with optimality guarantees, but they require comprehensive pre-modeling.
Which memory type, if skipped, most directly causes the agent to treat each session as brand new, breaking personalization?
- A Working
- B Episodic
- C Semantic
- D Sensory
Show answer
B — episodic. It provides cross-session continuity; without it, prior interactions can't be recalled.
Tree-of-Thought and Self-Ask are techniques used by which planning approach?
- A Symbolic planning
- B LLM-powered dynamic planning
- C STRIPS-based planning
- D Rule-based planning
Show answer
B. They leverage the LLM's reasoning to generate and explore solution paths dynamically, without rigid pre-programming.
Which architecture is best suited to real-time customer-service triage but risks hallucination without proper constraints?
- A Planning agent
- B Memory-Augmented agent
- C Autonomous Decision-Making agent
- D Multi-agent system
Show answer
C. Its emphasis on speed and immediacy is a strength for triage but a risk without safeguards; it also lacks strategic depth for complex tasks.
Part B · True or false
Symbolic planning adapts easily to novel, rapidly changing situations without pre-modeling.
Show answer
False. Symbolic planning offers precision and optimality but requires comprehensive pre-modeling — it's less adaptable to novel situations. That flexibility is the strength of LLM-powered dynamic planning.
The three foundational architectures are mutually exclusive and should not be combined.
Show answer
False. They're complementary building blocks — their true potential emerges through integration.
Semantic memory stores general factual knowledge, while episodic memory captures specific events and interactions.
Show answer
True. Semantic = concepts/facts (specs, regulations); episodic = timestamped experiences.
Planning agents are generally faster than Autonomous Decision-Making agents.
Show answer
False. Planning agents are typically slower — they analyze and design before execution — and need robust monitoring/recovery.
Hierarchical decomposition transforms high-level goals into a tree of smaller, concrete sub-goals.
Show answer
True. It mirrors how humans tackle large projects, making complex problems tractable.
Part C · Short answer
Name the three foundational cognitive architectures and each one's core strength in a phrase.
Show answer
Autonomous Decision-Making — rapid, reactive responses. Planning — strategic, multi-step goal orchestration. Memory-Augmented — continuity/personalization over time.
List the three memory types and when to query each.
Show answer
Working — active session, current turn (LLM context). Episodic — cross-session continuity, user references prior interaction (vector similarity search). Semantic — factual/domain knowledge not answerable from the session (keyword/semantic search over a knowledge base).
Contrast the two planning paradigms on flexibility vs. rigor.
Show answer
Symbolic (STRIPS/PDDL): rigorous, optimal, but needs full pre-modeling and is brittle to novelty. LLM-powered dynamic (ToT/Self-Ask): flexible and adaptive to novel situations, reasoning in natural language, but without formal optimality guarantees.
What is the key difference in focus between an Autonomous Decision-Making agent and a Planning agent?
Show answer
The decision-maker focuses on immediate, tactical, single decisions; the Planning agent focuses on strategic, long-term, multi-step coordinated processes with dynamic revision and cross-project memory.
What makes an Autonomous Decision-Making agent more than a simple reactive script?
Show answer
Enhanced perception adds situational awareness (system load, user tier, alerts), and its cognition scores candidate strategies across weighted axes to decide whether to resolve autonomously, escalate, or use guided resolution — strategic reasoning rather than fixed stimulus→response.