Curriculum Module 7
Advanced Prompt
Engineering for
Agentic Reasoning
Master the architecture of thought — designing prompts that don’t just instruct, but orchestrate multi-step reasoning, tool use, and autonomous decision-making.
Chain-of-Thought
ReAct Framework
Tool Calling
Self-Reflection
Multi-Agent
Memory Systems
Constraint Design
01 — Foundation
What is Agentic Reasoning?
Agentic reasoning is the capacity of a language model to decompose ambiguous goals into executable sub-tasks, maintain working memory across iterations, invoke external tools, evaluate intermediate results, and self-correct — all within a structured prompt architecture.
Goal Input
→
Decompose
→
Plan Steps
→
Execute + Observe
→
Reflect + Adapt
→
Final Answer
02 — Core Techniques
Six Pillars of Agentic Prompting
01
Chain-of-Thought (CoT)
Elicit stepwise reasoning by instructing the model to “think aloud.” Use zero-shot CoT triggers or few-shot exemplars to scaffold logical chains before committing to an answer.
02
ReAct Prompting
Interleave Reasoning traces with Action invocations. The model produces a thought, then a tool call, then observes the result — iterating until the task is resolved.
03
Reflective Self-Critique
Prompt the model to evaluate its own outputs against explicit criteria before finalizing. Use a Critic persona or structured rubric to surface and correct flawed reasoning.
04
Tree of Thoughts (ToT)
Explore multiple reasoning paths simultaneously. Score and prune branches using a heuristic evaluator, then backtrack to the most promising node — enabling non-linear problem solving.
05
Constrained Generation
Define output schemas, format contracts, and logical constraints in the system prompt. JSON mode, XML scaffolding, and regex grammars ensure deterministic structured outputs from probabilistic models.
06
Memory Architecture
Design working, episodic, and semantic memory layers within the context window. Use scratchpads, retrieval hooks, and summary buffers to maintain coherent long-horizon reasoning.
03 — Implementation
The ReAct Prompt Pattern
A canonical ReAct system prompt structures the model’s cognitive loop as an explicit, parseable trace — enabling reliable tool dispatch and graceful error recovery.
system_prompt.txt
# ── Agent Identity ────────────────────────────────── You are a reasoning agent with access to external tools. Solve the user's task using the following loop: THOUGHT: // Articulate your current understanding // and what sub-task to tackle next. ACTION: // Invoke exactly one tool per step. tool_name(param_a="value", param_b=42) OBSERVE: // The tool result is injected here automatically. # Repeat THOUGHT → ACTION → OBSERVE until resolved. # When confident, emit: ANSWER: // Your final, grounded response to the user. # ── Constraints ───────────────────────────────────── - Never fabricate tool results. - Always cite the OBSERVE block that supports your ANSWER. - Maximum 12 reasoning steps before requesting clarification.
04 — Strategy Map
Choosing the Right Technique
Not every problem demands the same cognitive architecture. Use this reference to match task properties to optimal prompting strategies.
| Task Type | Complexity | Recommended Technique | Key Prompt Element |
|---|---|---|---|
| Math / Logic | Medium | Zero-shot CoT | “Think step by step” |
| Research + Synthesis | High | ReAct + Tool Use | THOUGHT / ACTION / OBSERVE |
| Creative Writing | Low–Med | Few-shot Exemplars | 3–5 style demonstrations |
| Code Generation | High | CoT + Self-Critique | Inline test assertions |
| Long-Horizon Planning | Very High | Tree of Thoughts | Branch scoring + pruning |
| Structured Extraction | Low | Constrained Generation | JSON schema in system prompt |
| Multi-Agent Delegation | Very High | Orchestrator + Subagents | Role + capability manifests |
05 — Advanced Patterns
Multi-Agent Orchestration
The frontier of agentic prompting involves composing specialized sub-agents under a central orchestrator. Each agent receives a narrow, well-defined system prompt — maximizing coherence and minimizing context pollution.
orchestrator_pattern.py
# Orchestrator dispatches to specialist agents ORCHESTRATOR_PROMPT = """ You coordinate a team of specialist agents. Break the task into sub-problems and delegate: researcher → web search, fact retrieval analyst → data interpretation, reasoning writer → synthesis, structured output critic → quality review, error detection Emit delegation commands as JSON: { "delegate_to": "researcher", "task": "Find Q3 2024 revenue figures for NVDA", "output_format": "{ value: number, source: string }" } """ # Each specialist agent is a separate API call # with its own focused system prompt + task payload
06 — Design Principles
The Ten Laws of Agentic Prompts
Ⅰ. Explicit Persona
Define the agent’s identity, capabilities, and limitations in the system prompt. Ambiguous identity produces inconsistent behavior.
Ⅱ. Observable Reasoning
Force the model to externalize its reasoning before acting. Hidden reasoning leads to unverifiable, undebuggable outputs.
Ⅲ. Atomic Tool Calls
Each action must invoke exactly one tool with precise, typed parameters. Compound actions create unparseable outputs.
Ⅳ. Grounded Answers
Final responses must cite the observation(s) that justify them. Ungrounded answers are hallucination vectors.
Ⅴ. Bounded Iteration
Always specify a maximum step count. Unbounded loops consume tokens and stall — define an escape hatch.
Ⅵ. Failure Modes First
Enumerate what the agent should NOT do before describing what it should. Negative constraints are often more powerful.
Ⅶ. Schema Contracts
Define precise input/output schemas for every inter-agent message. Type ambiguity is the enemy of multi-agent reliability.
Ⅷ. Minimal Context
Include only the context needed for the current step. Bloated prompts dilute attention and degrade reasoning quality.
Ⅸ. Deterministic Defaults
Set temperature to 0 for reasoning and planning steps. Reserve higher temperatures for creative and generative stages.
Ⅹ. Eval-Driven Iteration
Treat every prompt as a hypothesis. Define measurable success criteria and run systematic evals before deploying agentic systems.
