0 / 1200 XP
NOVICE
PHASE 0
🧠
🧠 ASJPrompts & Studio · Claude Skill Engineering Bootcamp

AI Agent Engineering
Fundamentals

Not tools. Not frameworks. Fundamentals. Master the cognitive architecture, lifecycle, planning systems, memory models, and communication patterns that every production AI agent is built on.

10
Missions
1200
Total XP
5
Core Components
1
Architecture Challenge
WHY FUNDAMENTALS FIRST?

In 2026, Gartner forecasts 40% of enterprise applications will embed AI agents — up from under 5% in 2025. Multi-agent system inquiries surged 1,445% from Q1 2024 to Q2 2025. Yet most practitioners jump straight to frameworks without understanding why agents work the way they do.

This module fixes that. You will understand the cognitive primitives before you ever touch LangGraph, CrewAI, or any orchestration layer. The engineers who understand fundamentals are the ones who architect systems that survive production.

🧠 Cognitive Architecture 🔄 Agent Lifecycle 🗺️ Planning Systems 🔧 Tool Calling 🌐 Multi-Agent
Mission 1 — What is an AI Agent?

Russell and Norvig defined it in 1995: "Anything that perceives its environment through sensors and acts upon it through actuators." That definition is 30 years old. It still holds. The key word is acting — not responding.

M1
Agent Definition & Identity
3 concepts · +80 XP · Foundation of everything that follows
1
The Core Definition
What separates an agent from everything else
+25 XP

An AI agent is a software system that perceives its environment, reasons about what to do, and acts to achieve a goal — without being prompted at every step. Unlike a chatbot that waits for you, an agent pursues an objective.

The minimum viable agent has three things: (1) an LLM for reasoning, (2) a prompt defining its behavior and goal, and (3) an action space — the tools it can use to act in the world. Remove any of these and you don't have an agent.
🎯
Goal-Driven
Given an objective, not a command. The agent determines how to achieve it.
🔁
Autonomous Loop
Runs continuously: perceive → reason → act → observe → repeat until done.
Action-Capable
Takes real actions in the world — search, write, execute, call APIs, delegate.
2
Types of Agency
From simple reflex to learning systems
+30 XP

Not all agents are equal. The field classifies agents by how they make decisions:

Agent TypeDecision BasisMemoryLLM Era?
Simple ReflexCondition → Action rules✗ None
Model-Based ReflexInternal world modelLimitedPartial
Goal-BasedActions aligned to goalsSession
Utility-BasedBest expected outcomeSession
Learning AgentImproves from feedback✓ Persistent✓ Modern
2026 Reality: Most production LLM agents are goal-based with elements of utility reasoning. True persistent learning agents (self-improving across sessions) are still emerging — agents like Reflexion and Mem0-based systems approach this.
3
Agenticness: The Spectrum
How autonomous is autonomous?
+25 XP

Gartner named Agentic AI the #1 strategic technology trend for 2025. But "agency" is a spectrum, not a binary. The CoALA framework defines agenticness across four dimensions:

🎯
Goal Complexity
Range and difficulty of tasks — reliability, speed, safety tradeoffs.
🌍
Environmental Complexity
Multi-stakeholder, long-horizon, cross-domain contexts.
🔄
Adaptability
Responding to novel, unexpected circumstances in real time.
👁️
Supervision Level
How much direct human oversight is required per decision.
Key insight: High agenticness = high goal complexity + high adaptability + low supervision needed. Most enterprise agents in 2026 sit mid-spectrum — fully autonomous for bounded tasks, human-in-the-loop for high-stakes decisions.
🤖
Agent Definer
Mission 1 complete — you can define an AI agent from first principles
+80 XP
Mission 2 — Agent vs Assistant

Same underlying technology. Fundamentally different operational mode. Autonomy is the dividing line — assistants wait for prompts and reset with each interaction; agents work toward goals, maintain context, and adapt when conditions change.

M2
Autonomy is the Dividing Line
Comparison · Use cases · When to choose which · +80 XP
DimensionAI AssistantAI Agent
TriggerUser promptGoal assignment
Interaction modeReactiveProactive
Session memoryResets each timePersists across sessions
Task scopeSingle-turn, definedMulti-step, open-ended
Tool useLimited / on requestAutonomous tool selection
Planning✗ Doesn't plan✓ Plans full strategy
Error recoveryUser must redirectSelf-corrects autonomously
Multi-systemSingle conversationAPIs, agents, databases
Human input neededEvery stepGoal + boundaries only
Best forQuick tasks, answersEnd-to-end workflows
📋
The HITL Middle Ground
Human-in-the-loop agents — the 2026 production standard
+40 XP

The assistant vs agent binary is a simplification. The 2026 production standard is Human-in-the-Loop (HITL) agents — systems that act autonomously within guardrails but escalate to humans for high-stakes decisions.

HITL agents take actions based on context and goals (like a full agent) while still responding to human inputs where necessary. This balance between oversight and automation is why enterprises are deploying HITL patterns over fully autonomous ones in regulated industries like pharma and finance.
🔵
Use an Assistant when...
You need content, analysis, summaries, or answers. Task ends after your next read.
🟣
Use an Agent when...
You need tasks completed across multiple systems without manual handoffs between steps.
🟢
Combine both when...
The assistant clarifies goals and provides feedback; agent handles multi-step work in background.
🏭
Real-World Contrast
Same scenario — assistant vs agent outcome
+40 XP

Scenario: "Migrate customer data from CRM A to CRM B."

🔵
ASSISTANT:Asks "How should I do this?" Waits for manual step-by-step direction. Produces output only when prompted. Stops if you stop asking.
🟣
AGENT:Reads CRM A schema → maps fields → validates mismatches → exports batch → cleans nulls → imports to CRM B → verifies row count → generates migration report. Flags anomalies. Done.
The real difference: The assistant makes recommendations. The agent evaluates solutions and independently executes. Where an assistant makes you the executor, an agent makes the LLM the executor.
Mission 3 — Agent Components

Every production agent is built on five cognitive primitives. Master these and you can understand, audit, or design any agent system regardless of the framework wrapping them.

M3
The 5 Primitives: Goals · Memory · Tools · Reasoning · Actions
Deep dive into each layer · +120 XP
🎯
Goals
The agent's objective state — what it is trying to achieve
+20 XP

Goals define the agent's purpose and termination condition. Without a clear goal, an agent cannot decide which actions are useful. Goals can be:

📌
Terminal Goals
Complete a specific task and stop. "Summarize all 50 papers and output a report."
🔄
Continuous Goals
Ongoing monitoring or maintenance. "Keep the CI pipeline green at all times."
📊
Utility Goals
Maximize a metric. "Generate the highest-ICER value within budget constraints."
Goal engineering is the most underrated skill in agent design. An ambiguous goal produces an agent that wanders. A precise goal with clear success criteria produces an agent that executes. The system prompt is where goals live — it is not boilerplate; it is the agent's mandate.
🧠
Memory
Four memory types from the CoALA framework (Princeton, 2023)
+25 XP

The CoALA framework formalizes four memory types. IBM, MongoDB, LangChain, Letta, and Mem0 all use versions of this model in their agent documentation:

In-Context (Working)
Immediate context window — the agent's RAM. Active during the current reasoning step only.
Current conversation, tool outputs, task state
📖
Episodic
Past experiences with temporal details. Enables case-based reasoning — "last time X happened, I did Y."
Vector DB logs of past tasks and outcomes
📚
Semantic
Factual knowledge independent of specific events. Domain expertise, user profiles, product specs.
RAG knowledge base, structured company data
⚙️
Procedural
Rules, skills, and learned strategies for executing tasks. How the agent knows to do things.
SKILL.md files, AGENTS.md, system prompt rules
December 2025 research ("Memory in the Age of AI Agents", arXiv:2512.13564) notes the field is fragmenting as episodic memory research accelerates. Episodic memory — per a 2025 position paper — requires: long-term storage, explicit reasoning, single-shot learning, instance-specific details, and contextual binding.
🔧
Tools
The agent's hands — external capabilities beyond the LLM
+20 XP

LLMs are isolated reasoning engines. Tools are what connect them to the world. Without tools, an agent cannot act — it can only think. Tools come in four categories:

🔍
Retrieval Tools
Search, web browsing, vector DB queries, document retrieval. Grounds the agent in real data.
READ
Execution Tools
Code execution, bash, system commands. Agent can compute, not just reason about computing.
RUN
🔌
API Tools
REST calls, database writes, CRM updates, calendar events. The agent affects external systems.
CALL
🤖
Agent Tools
Spawning subagents, delegating tasks, calling specialist agents. Multi-agent composition.
DELEGATE
💡
Reasoning
The cognitive engine — how agents think before acting
+25 XP

The LLM is not the agent — it is the reasoning engine inside the agent. Reasoning is how the agent interprets inputs, evaluates options, and selects actions. Three core reasoning modes:

1
Deductive Reasoning
Applies known rules to reach conclusions. "IF inventory < threshold AND lead time = 2 weeks THEN trigger reorder." Deterministic, fast.
2
Inductive Reasoning
Generalizes from observations to patterns. Learns that "queries arriving after 3pm tend to be lower priority." Experience-based.
3
Abductive Reasoning
Best explanation for incomplete evidence. "The model output is inconsistent — most likely the context window was truncated." Hypothesis-driven.
Modern LLM agents blend all three. Chain-of-Thought prompting unlocks explicit step-by-step deduction. ReAct interleaves inductive observation with action. Reflexion enables abductive hypothesis correction from failure.
Actions
What the agent actually does in the world
+30 XP

Actions are the output of reasoning. Every action an agent takes falls into one of these categories:

📝
Generation
Producing text, code, reports, structured data. The LLM's native output mode.
🔍
Information Retrieval
Querying memory, databases, search engines, or other agents for data.
🌐
Environment Interaction
Writing files, calling APIs, executing code, clicking UI elements, sending messages.
🤖
Agent Delegation
Spawning subagents or handing off to specialists with structured briefs.
🛑
Termination
Deciding the goal is achieved (or cannot be achieved) and stopping the loop.
Termination is an action. An agent that cannot decide to stop is dangerous in production — it loops infinitely, burning compute and potentially causing side effects. Every production agent needs explicit termination conditions.
🧩
Component Architect
All 5 agent primitives mastered — you can audit any agent system
+120 XP
Mission 4 — Agent Lifecycle

The agent loop is the heartbeat of autonomous AI. Every major AI company — Anthropic, OpenAI, Google, Microsoft — has converged on the same five-stage cycle, despite building very different products around it.

M4
Perceive → Reason → Plan → Act → Observe
The PRAL loop · Production constraints · +90 XP
THE AGENT LOOP — RUNS UNTIL GOAL MET OR TERMINATION
👁️
PERCEIVE
Receive inputs, context, observations from environment
💡
REASON
LLM processes context, evaluates options, forms intent
🗺️
PLAN
Decompose goal into steps, select tools, sequence actions
ACT
Execute tool calls, generate outputs, delegate subtasks
🔍
OBSERVE
Receive results, update state, decide: done or loop again?
↑ Loop continues back to PERCEIVE until termination condition met ↑
📐
Each Stage Explained
What actually happens inside each loop phase
+45 XP
1
PERCEIVE — Input Processing
The agent receives raw inputs: user messages, tool results, sensor data, API responses, or previous loop observations. The perception layer converts these into structured context the reasoning engine can process. This includes context window management, conversation state tracking, and input validation.
2
REASON — Cognitive Processing
The LLM interprets the context against its goal and decides what to do next. This is not just "generate text" — it is evaluating whether current state satisfies the goal, which tools are relevant, and what the optimal next action is. Chain-of-Thought, ReAct, and planning patterns all happen here.
3
PLAN — Decomposition
For complex goals, the agent breaks the objective into sub-tasks, sequences them, identifies dependencies, and selects tools for each step. Simple agents skip explicit planning (reactive ReAct); complex agents generate full plans before executing (Plan-and-Execute).
4
ACT — Execution
The agent executes the chosen action: calling a tool, generating output, delegating to a subagent, or writing to memory. This produces a result that is passed to the next stage. Agents consume approximately 4× more tokens than standard chat interactions here.
5
OBSERVE — State Update
The agent receives the action result, updates its internal state, and evaluates termination: is the goal achieved? Has a stopping condition been hit? If not, it loops back to PERCEIVE with new context. The loop continues until the goal is met or a max iteration/timeout is reached.
The ReAct breakthrough: Princeton/Google Research (2022) showed that interleaving reasoning and action in a single loop produced a 34% improvement on ALFWorld and 10% on WebShop versus single-pass responses. The loop is not overhead — it is the source of performance.
⚠️
Production Constraints
What engineers actually wrestle with in real deployments
+45 XP
More tokens vs chat
15×
Multi-agent token cost
5
Stages per loop
Loops (without stop)
Agent loops need guardrails: Max iteration limits. Timeout conditions. Cost budgets. Graceful degradation when tools fail. Without these, a production agent will loop indefinitely, burning compute and producing side effects. Every loop must have an exit.
Observability is the other constraint: tracing every reasoning step, tool call, and decision across an iterative loop is fundamentally harder than tracing a single LLM call. Engineers call debugging buried errors in long agent traces "AI archaeology."
Mission 5 — Planning Systems

Planning is how agents decompose complex goals into executable steps. The right planning strategy determines whether your agent solves hard problems reliably or fails unpredictably under complexity.

M5
CoT → ReAct → Tree of Thoughts → Plan-and-Execute
4 core planning paradigms with benchmark data · +100 XP
🔗
Chain-of-Thought (CoT)
The foundation all other patterns build on. Forces the model to reason step-by-step before answering. Dramatically reduces errors on multi-step problems. Zero-shot CoT ("let's think step by step") works surprisingly well. Self-Consistent CoT generates multiple reasoning paths and picks the majority answer.
→ Use for: Mathematical reasoning, logical deduction, structured analysis. Best on well-defined problems with clear steps.
ReAct (Reasoning + Acting)
The default agent pattern in LangChain and LangGraph. Alternates reasoning traces ("Thought:") with tool actions ("Action:") and results ("Observation:") in a live loop. Each observation updates the reasoning context — the agent adapts dynamically. Outperformed standalone CoT by 34% on ALFWorld (ICLR 2023). Best balance of adaptability and reliability.
→ Use for: Dynamic real-time tasks where intermediate results change the plan. Search, data gathering, multi-step workflows.
🌳
Tree of Thoughts (ToT)
Explores multiple solution paths simultaneously using tree search. At each step, the agent generates N candidate actions, evaluates them, and keeps the most promising branches. Solved 74% of Game of 24 math tasks with GPT-4 vs just 4% with standard CoT. Tradeoff: 10–100× more compute. Not for real-time tasks.
→ Use for: Complex problems where linear reasoning fails. Creative problem-solving, optimization, scenarios with many viable paths.
📋
Plan-and-Execute
Decouples planning from execution. Phase 1: use an expensive model to generate a complete plan upfront. Phase 2: use cheaper, faster models to execute each step. Benchmarks show up to 92% task completion with 3.6× speedup over sequential ReAct. Re-planner evaluates results after execution and adjusts the plan if needed. Now the production standard for long-horizon tasks.
→ Use for: Multi-step workflows with known structure. Research pipelines, document processing, multi-system integration.
📊
Pattern Selection Guide
Which planning system to use when
+50 XP
PatternSpeedCostFlexibilityBest When
CoTFastLowMediumStructured, predictable problems
ReActMediumMediumHighDynamic tasks, real-time data needed
Tree of ThoughtsSlowHighVery HighComplex optimization, creative tasks
Plan-and-ExecuteFast execMediumMediumMulti-step workflows, known structure
ReflexionMediumMediumHighTasks requiring iterative refinement
2026 production reality: Most systems combine patterns. Plan-and-Execute + ReAct is common: plan the strategy with an expensive model, then execute each step with ReAct's adaptive loop. This gives you predictable cost + dynamic adaptation.
🧮
Planning Prompt Anatomy
What a real ReAct planning trace looks like
+50 XP

A ReAct trace in production looks like this — Thought/Action/Observation repeating until done:

T1
Thought:I need to find the ICER threshold for the UK market. Let me search NICE guidelines.
A1
Action:search_tool("NICE ICER threshold 2025 UK health technology assessment")
O1
Observation:NICE standard threshold: £20,000–£30,000/QALY. End-of-life modifier: up to £50,000/QALY.
T2
Thought:Got UK threshold. Now I need the drug's modeled ICER from the submitted dossier.
A2
Action:retrieve_document("CardiozemX_HTA_dossier.pdf", section="cost_effectiveness")
O2
Observation:Base case ICER: £24,800/QALY. PSA 95th percentile: £38,200/QALY.
T3
Thought:Base case is within threshold. PSA tail exceeds it. Recommendation: likely positive with risk-sharing. Task complete.
DONE:Goal achieved. Output generated. Loop terminates.
Mission 6 — Tool Calling

Tool calling transforms an LLM from a passive text generator into an active agent that interacts with external systems. It is the single capability that makes agents possible. Without it, you have a very smart calculator that cannot touch the world.

M6
The 6-Step Tool Calling Loop
JSON schema · Tool selection logic · MCP protocol · +90 XP
🔧
How Tool Calling Works
The 6-step modern production loop
+45 XP

In 2026 production environments with dynamic discovery, tool calling is a 6-step process — evolved from the simpler 5-step early documentation:

0
Tool Discovery (Step 0)
The system queries a Tool Registry (via MCP or vector store) based on the user's intent. This prevents context window saturation — the agent sees only relevant tool definitions, not all 200+ available tools.
1
Tool Definition
The LLM receives specific JSON Schema definitions for discovered tools. Each definition includes: name, description, parameters (with types and required fields). The LLM uses this to understand what each tool does and what input it expects.
2
User Prompt
The user provides a request that requires external action. The system prompt + tool definitions + user message form the complete context.
3
LLM Prediction
The model analyzes the prompt against available tool definitions and outputs a structured JSON payload — the "Tool Call." This contains the tool name and all required arguments. The model outputs special tokens that signal tool use rather than regular text.
4
Execution
The application layer intercepts the tool call JSON, validates arguments, authenticates, and executes the actual function. The LLM does not execute code — it outputs structured intent; the application executes it.
5
Result Return
The tool result is returned to the LLM as a new context entry. The model processes the result and decides: answer the user, call another tool, or loop again. This is where agent behavior emerges from single tool calls.
The real bottleneck is not the LLM's reasoning — it's the integration plumbing. Authenticating across APIs, mapping parameter formats, handling errors gracefully, rate limiting. This is why Model Context Protocol (MCP) exists: standardized plumbing so engineers focus on agent logic.
📄
Tool Schema Anatomy
What a tool definition actually looks like
+45 XP
TOOL DEFINITION (JSON Schema)
{
"name": "search_clinical_trials",
"description": "Search ClinicalTrials.gov for trials by drug name, phase, condition, or sponsor.",
"parameters": {
"drug_name": {"type": "string", "description": "INN name of the drug"},
"phase": {"type": "string", "enum": ["I","II","III","IV"], "description": "Trial phase"},
"required": ["drug_name"]
}
// LLM reads this and decides when to call and with what args
Description quality is everything. The LLM decides which tool to call based entirely on the description. A vague description = wrong tool selection. A precise description that explains exactly what the tool does, when to use it, and what format the output takes = correct tool selection. Write tool descriptions like you're writing prompts.
MCP (Model Context Protocol) — created by Anthropic, open-sourced November 2024, donated to the Linux Foundation December 2025 — standardizes how agents connect to external tools. Think of it as HTTP for agent-to-tool connections. It eliminates custom integration code for every new tool.
Mission 7 — Reflection

Reflection is how agents evaluate and improve their own outputs before finalizing them. Andrew Ng named it one of the four core design patterns for agentic AI. It's one of the simplest ways to make an agent significantly more reliable — and measurably so.

M7
Self-Critique · Reflexion · Adaptive Reflection
+14–20% accuracy improvement per session · +80 XP
🪞
The Reflection Loop
Generate → Critique → Refine → Evaluate
+40 XP

Without reflection, agents repeat the same errors. With it, they get measurably better — 14–20% accuracy improvement in a single session. The reflection loop has four stages:

1
GENERATE — Initial Output
The agent produces its first response or action. This is the first-pass output before any self-review.
2
C
CRITIQUE — Self-Evaluation
The agent shifts into reviewer mode: checking for logical gaps, inconsistencies, missing requirements, factual errors, incomplete coverage. This perspective shift alone surfaces issues invisible in the initial pass.
3
REFINE — Improvement
Based on critique, the agent revises the output. Iterates until quality thresholds are met. The Reflexion pattern stores what went wrong as explicit memory for future episodes.
4
EVALUATE — Quality Gate
External criteria check — not just self-assessment. Pass/fail thresholds, measurable standards, business-aligned metrics. Without evaluation grounded in external criteria, reflection optimizes appearance rather than correctness.
Reflexion (2023) achieved 91% pass@1 on HumanEval coding benchmarks — surpassing GPT-4's prior state-of-the-art of 80%. The improvement came entirely from the self-reflection loop, not from a stronger underlying model. Same model, better architecture.
⚠️
Reflection Anti-Patterns
Where reflection fails in production
+40 XP
Self-consistency trap (EMNLP 2025): LLMs can generate plausible but incorrect content with high internal self-consistency. A model may confidently defend a wrong answer through multiple reflection rounds. Ground reflection in external tools and data sources, not just the model's internal judgment.
Sycophantic reflection: Models may agree with their own outputs rather than genuinely critiquing them — especially when the initial output sounds authoritative. Use different models or different temperatures for generation vs. critique to increase diversity.
Over-reflection: Research shows reflection can actually hurt performance on tasks where the initial response is already highly accurate. Selective reflection — triggered only when task difficulty warrants it — is more efficient and avoids introducing new errors.
Cost amplification: Each reflection round multiplies compute costs. In production, reflection must be paired with evaluation thresholds that stop the loop when quality is sufficient, not run indefinitely.
Production rule: Reflection is never used alone. It's combined with: tool use (to verify facts), human-in-the-loop controls (for high-risk outputs), evaluation patterns (to enforce standards), and orchestration logic (to decide when reflection is sufficient vs. when to escalate).
Mission 8 — Agent Communication

Agents don't just talk to users — they talk to other agents, tools, and systems. The protocols and patterns governing these communications are the connective tissue of every multi-agent system in production.

M8
MCP · A2A · Orchestrator-Subagent · Message Structure
2026 production communication patterns · +90 XP
📡
The Two Protocols of 2026
MCP and A2A — the HTTP of the agentic web
+45 XP

Just as HTTP enabled any browser to talk to any server, two open standards now do the same for AI agents — making them composable and interoperable across platforms and vendors:

🔌
MCP (Model Context Protocol)
Created by Anthropic. Open-sourced Nov 2024. Donated to Linux Foundation Dec 2025. Standardizes how agents connect to external tools, databases, and APIs. Eliminates custom integration code. De-facto standard in 2026.
AGENT ↔ TOOL
🤝
A2A (Agent-to-Agent)
Created by Google. Defines how agents from different vendors communicate directly. Enables cross-platform multi-agent collaboration. Complementary to MCP. Allows Anthropic agents to talk to OpenAI agents to custom Python agents.
AGENT ↔ AGENT
MCP handles agent-to-tool connections. A2A handles agent-to-agent communication. Together they form the plumbing of the agentic web. MCP without A2A = agents that can use tools but can't coordinate with each other. A2A without MCP = agents that can talk but can't act.
📨
Structured Agent Briefs
The P2 pattern — how orchestrators communicate with subagents
+45 XP

Every surviving production multi-agent deployment (validated across 2025–2026) uses the P2 prompt pattern — a structured contract between orchestrator and subagent. Free-form delegations are a documented failure mode.

P2 STRUCTURED BRIEF — ORCHESTRATOR → SUBAGENT
1
OBJECTIVE: "Retrieve all Phase III trials for GLP-1 agonists in cardiovascular outcomes from ClinicalTrials.gov."
2
OUTPUT FORMAT: "Return JSON array: [{nct_id, drug, phase, endpoints, n_patients, status}]"
3
TOOLS: "Use search_clinicaltrials tool only. Do not use web search."
4
BOUNDARIES: "Max 50 results. Exclude terminated trials. Date range: 2020–2025."
RETURN:Summary string, not full transcript. Forward directly to next agent.
Rule 4 (often missed): When the supervisor's only job is to deliver the subagent's output — forward it directly. ~50% of performance gain in supervisor-worker systems comes from this single change. Inlining full transcripts pollutes context and burns tokens at 15× the rate.
Mission 9 — Agent Patterns

Three archetypal agent personas appear across every industry: the researcher who gathers intelligence, the analyst who synthesizes it, and the planner who converts it into action. Master these three and you can staff any agent team.

M9
Research Agent · Analyst Agent · Planner Agent
Identity · Tools · Reasoning mode · Output contracts · +100 XP
🔍
Research Agent
Intelligence gatherer — broad, deep, sourced
+30 XP
RESEARCH AGENT PROFILE
GATHERER
🎯
Goal
Systematically gather, filter, and structure information from multiple sources on a defined topic. Produces sourced intelligence, not opinions.
🔧
Primary Tools
Web search, PubMed/ClinicalTrials.gov/SEC EDGAR APIs, document retrieval, web scraping, vector DB queries. Reads broadly across sources.
💡
Reasoning Mode
Breadth-first then depth. ReAct for dynamic search — each result informs the next query. Stops when coverage threshold is met or no new information is being discovered.
📤
Output Contract
Structured JSON or markdown with: source citations, extracted key facts, confidence levels, coverage gaps. Never synthesizes — that is the analyst's job.
🧠
Memory Needs
Episodic memory of past searches (avoid re-querying same sources). Semantic memory of domain vocabulary. Working memory for current search state.
📊
Analyst Agent
Intelligence synthesizer — pattern finder, insight generator
+35 XP
ANALYST AGENT PROFILE
SYNTHESIZER
🎯
Goal
Transform raw research outputs into structured insights, comparisons, and evidence-based conclusions. Identifies patterns the researcher surfaced data about.
🔧
Primary Tools
Code execution (statistical analysis), data transformation, chart generation, contradiction detection, NMA (network meta-analysis) tools, structured output validators.
💡
Reasoning Mode
Structured comparison with reflection. Evaluates evidence quality, weights sources, identifies contradictions, applies domain frameworks (GRADE, PICOT, ICER thresholds). Uses Reflexion to self-check conclusions.
📤
Output Contract
Evidence tables, key findings with confidence ratings, identified gaps, structured comparisons. Never recommends actions — that is the planner's job.
🧠
Memory Needs
Semantic memory of domain frameworks and benchmarks. Working memory for current analysis state. Episodic memory of past analytical conclusions on similar topics.
🗺️
Planner Agent
Decision converter — turns insights into executable strategy
+35 XP
PLANNER AGENT PROFILE
STRATEGIST
🎯
Goal
Convert analyst insights into prioritized, time-sequenced, executable action plans with resource assignments and success metrics defined.
🔧
Primary Tools
Project management APIs, calendar/scheduling, task creation, dependency mapping, risk assessment tools, notification systems. Executes and coordinates rather than researches.
💡
Reasoning Mode
Plan-and-Execute. Creates full strategy upfront from analyst inputs, then delegates execution steps to executor agents or humans. Re-plans when obstacles surface. Manages dependencies and critical path.
📤
Output Contract
Structured plans with: objectives, time-sequenced tasks, assigned owners (agents or humans), dependencies, KPIs, risk flags, contingency triggers.
🧠
Memory Needs
Procedural memory of planning templates and past successful plan structures. Working memory for current plan state. Episodic memory of plan execution outcomes for retrospectives.
Pattern synergy: Research Agent → Analyst Agent → Planner Agent is the fundamental pipeline for any intelligence-driven workflow. The researcher surfaces facts. The analyst produces insight. The planner produces action. In production, these run in sequence or in parallel depending on the task — this is the backbone of an HEOR multi-agent system, a CI pipeline, or a market access engine.
Mission 10 — Multi-Agent Concepts

Single agents hit ceilings. Multi-agent systems break through them — by parallelizing work, isolating domains, and enabling specialist agents to do what generalists can't. Gartner logged a 1,445% surge in multi-agent inquiries from Q1 2024 to Q2 2025.

M10
Orchestrator-Worker · Topologies · 3 Patterns That Survived 2026
Production-validated patterns only · +110 XP
🌐
Why Multi-Agent?
When single agents fail and multi-agent succeeds
+35 XP

Multi-agent systems break through single-agent ceilings in five scenarios:

⏱️
Parallelism
Tasks that can run simultaneously. Research 5 drugs in parallel → 5× faster than sequential.
🏛️
Domain Isolation
Compliance boundaries. One agent handles PII, another handles public data — never mixing.
🎓
Specialization
A statistics agent calibrated for HEOR outperforms a generalist agent on every ICER calculation.
🔄
AI-to-AI Validation
One agent generates, a second independently validates. Catches errors before human review (ISPOR 2025).
📏
Context Limits
Tasks too large for one context window — decompose across agents, each handling a bounded scope.
ISPOR Europe 2025 explicitly documented that multi-agent systems in HEOR enable AI-to-AI validation and correction before human review, "significantly improving output reliability while reducing review time and costs." This is now an accepted production pattern in health economics.
🏗️
3 Topologies That Survived 2026
What actually works in production — validated across deployments
+40 XP

Peer-collaboration multi-agent systems failed production in 2026. Only three patterns survived — validated across real deployments including $75K/day mission-critical systems:

👑
1. Supervisor-Worker (Default 2026)
A single orchestrator owns full conversation context, spawning ephemeral isolated subagents that return compressed summaries. No peer-to-peer communication. Anthropic's "brain/hands" architecture, OpenAI's Agents SDK, LangChain's supervisor pattern — all converged here. Five major vendors implemented this in 2025–2026.
→ Use for: Domain isolation, compliance boundaries, parallel independent research queries
🌊
2. Sequential Pipeline (Handoff)
Agent A completes its task and hands a structured payload to Agent B, which hands to Agent C. OpenAI's Agents SDK ships a first-class handoff primitive with full replay telemetry. The implicit topology is orchestrator-worker but chained — each agent's output is the next agent's input.
→ Use for: Research → Analysis → Planning pipelines. Dossier generation. Multi-stage document processing.
3. Fan-Out / Fan-In (Parallel)
Orchestrator dispatches N workers concurrently for independent subtasks, then merges outputs via an aggregator. Best when sub-tasks have no interdependencies. A supervisor routes to specialist agents simultaneously — research 5 competitors in parallel, merge findings, synthesize once.
→ Use for: Competitive intelligence across multiple products. Parallel evidence synthesis. Multi-country market access.
Free mesh failed: Systems where any agent could communicate with any other agent created uncontrollable information loops, context pollution, and debugging nightmares. Only survived as a controlled subroutine inside a supervisor, never as the top-level architecture.
⚙️
Orchestration Layer Responsibilities
What the coordinator must manage
+35 XP
1
Dependency Graph Management
Track which tasks must complete before others can start. Enforce sequencing. Detect circular dependencies before spawning.
2
Communication Protocol Control
Route messages correctly between agents. Prevent context pollution. Enforce the P2 brief structure for all delegations.
3
Shared Memory & State Persistence
Maintain consistent shared state. Decide what each subagent can see. Prevent agents from overwriting each other's memory.
4
Agent Health & Retry Policy
Monitor subagent status. Implement retry with backoff. Escalate to human if a subagent fails beyond tolerance.
5
Intent Routing
Classify incoming tasks and route to the correct specialist agent. Taxonomy of agent capabilities must be maintained and queried by the orchestrator.
🌐
Multi-Agent Architect
Mission 10 complete — you understand multi-agent systems at the production level
+110 XP
Final Assessment

Design an HEOR Agent architecture. No building. No code. Pure architecture thinking. This is what separates engineers who understand fundamentals from those who just copy framework tutorials.

📋
Architecture Challenge: HEOR Agent System
Quiz (10 questions, 80% to pass) + Architecture Design · +200 XP on pass
THE ARCHITECTURE CHALLENGE

You are designing a multi-agent HEOR intelligence system for a pharmaceutical company. The system must: (1) gather evidence from published literature and clinical databases, (2) synthesize cost-effectiveness data against NICE/IQWiG/HAS thresholds, (3) produce a structured HTA dossier narrative, and (4) flag contradictions and evidence gaps for human review.

No building required. Map the architecture: Which agents? What roles? What topology? What memory does each agent need? What tools does each use? What are the communication contracts between them?

REFERENCE ARCHITECTURE (Study Before Quiz)
🔍 Evidence Scout Agent
Systematic literature retrieval from PubMed, EMBASE, ClinicalTrials.gov. Filters by PICOS criteria. Returns structured citation JSON.
Tools: pubmed_search, clinicaltrials_api, document_retriever
📊 CE Analyst Agent
Extracts ICER values, QALY estimates, utility scores. Compares against NICE (£20-30K), IQWiG (added benefit), HAS (ASMR) thresholds.
Tools: data_extractor, threshold_lookup, markov_model_reader
🔗 NMA Synthesis Agent
Runs network meta-analysis across indirect comparators. Produces comparative efficacy estimates with credible intervals for the dossier.
Tools: nma_executor, forest_plot_generator, consistency_checker
✍️ Dossier Writer Agent
Synthesizes analyst outputs into structured HTA dossier sections. Follows CTD/eCTD format. Flags evidence gaps and contradictions inline.
Tools: template_engine, citation_formatter, gap_detector
⚖️ QA Validator Agent
Independent validation pass before human review. Checks source traceability, GRADE evidence quality, logical consistency. Returns structured review report.
Tools: grade_assessor, citation_verifier, consistency_auditor
👑 HEOR Orchestrator
Supervisor-worker topology. Manages dependency graph: Scout → CE Analyst + NMA (parallel) → Writer → QA. Routes human escalations.
Topology: Fan-out (CE+NMA), Sequential (→Writer→QA), HITL escalation
🏆 Your Certificate

Complete all missions and pass the assessment with 80%+ to unlock your sovereign certificate of completion.

🔒
Certificate Locked
Complete the requirements below to unlock your Certificate of Completion for AI Agent Engineering Fundamentals.
🎯
Assessment Score: 80% or above required on the 10-question quiz
Missions Completed: Complete missions across all 10 modules to earn XP
🧠
Fundamentals Mastered: Goals, Memory, Tools, Reasoning, Actions, Planning, Reflection, Communication, Patterns, Multi-Agent