AI Agents Full Course 2026: Master Agentic AI (2 Hours)

Author: Nick Saraev · URL: watch (2h)

Summary

Comprehensive 2-hour course on AI agents and agentic AI in 2026. Covers the core agent workflow loop, multi-agent orchestration, advanced prompting techniques, and practical implementation across platforms (Codex, Claude Code, Antigravity). Demonstrates how to build systems that parallelize work across multiple agent instances for economically valuable tasks like lead generation, form filling, and web automation.

Key Points

Core Agent Architecture

• Agent Loop: Observe → Think → Act → Repeat

  • Observe: Agent reads all context (files, tool calls, system prompts, research, multimodal data)
  • Think: Dedicated reasoning step where model plans its approach (visible in Claude/Gemini reasoning tabs)
  • Act: Calls tools, edits files, runs CLI commands
  • Loop: Results feed back into observe step, growing context each iteration
  • Definition of Done: Critical constraint/specifications that signal task completion—often missing from novice prompts

• Agents vs Chatbots: Agents = LLM + Tools + Reasoning Loop + Memory. The LLM alone is just the reasoning engine; infrastructure around it enables agentic behavior.

Platform Comparison (2026)

Platform	Model	Strengths	Weaknesses
Claude Code	Claude	Most interpretable reasoning, great for orchestration, consistent quality	Slower (unless fast mode), weaker at frontend design
Antigravity	Gemini	Best at frontend/design, superior multimodal (video understanding), fast output	Least interpretable, inconsistent quality
Codex	GPT 5.4	Best at backend programming, math, test-driven development, large ecosystem	Less interpretable than Claude

Note: Differences are minor (2-5%)—all models trained on internet-scale data. Choose based on specific needs.

Foundational Prompting Techniques

1. Self-Modifying System Prompts (agents.md/claude.md/gemini.md)

   • Create a file at project root that prepends to every conversation
   • Include meta-prompt that instructs agent to update file with new rules when corrected
   • Rules accumulate across sessions, reducing errors over time
   • Structure: “When user corrects you, immediately append a new rule to learned rules section”
   • Global rules apply to all projects; local rules apply per-project

2. Agent Skills

   • Standardized workflows defined as .md files with YAML frontmatter (name, description, tools)
   • Makes flexible LLMs deterministic for repeatable tasks
   • Available on all major platforms (Codex, Gemini, Claude)

Advanced Multi-Agent Patterns

3. Multi-Agent MCP Orchestration

   • Use one model as orchestrator/manager (typically Claude for interpretability)
   • Delegate subtasks to specialized models: Gemini for frontend, Codex for backend/testing
   • Requires API keys for each platform
   • Parallelizes work while getting percentage-point quality gains
   • Higher cost (no platform subsidization via API)

4. Video-to-Action Pipelines

   • Leverages Gemini’s native video understanding via API
   • Flow: YouTube URL → Claude receives URL → calls Gemini API → Gemini extracts step-by-step instructions → Claude executes with tools
   • Enables agents to learn from same medium as humans (tutorials, demos)
   • Example: Replicating Blender donut tutorial or N8N workflow from video

5. Stochastic Multi-Agent Consensus

   • Exploits model stochasticity (randomness) to traverse larger solution space
   • Spawn N agents with slight framing variations for same problem
   • Parent agent aggregates results: identifies consensus items, divergent items, and outliers
   • Use case: Ideation, strategic analysis, filtering hallucinations
   • Example: 10 agents analyzing TikTok growth strategy surfaced both consensus (hook reformatting) and outlier ideas (paid spark ads)

6. Agent Chat Rooms

   • Agents with different personalities debate in shared chat.json
   • Round-robin turns: Systems thinker → Pragmatist → Edge case finder → User advocate → Contrarian
   • Challenges assumptions, sharpens ideas, catches errors through disagreement
   • Produces higher-quality, more nuanced outputs than parallel independent agents

7. Sub-Agent Verification Loops

   • Implementer agent writes first draft → passes output (not reasoning) to reviewer agent with fresh context
   • Reviewer has no sunk cost bias, catches issues implementer missed
   • Resolver agent fixes issues if found
   • Similar to academic peer review—objective evaluation improves quality

Prompt Engineering Patterns

8. Prompt Contracts

   • Before implementing, agent generates structured contract with: Goal, Constraints, Format, Failure conditions
   • Forces clarification of vague requests (e.g., “beautiful site” → “single-page, linear white aesthetic, <500 lines HTML, no Bootstrap look”)
   • Similar to scoping in freelance work—navigates line between too vague and too restrictive
   • Can be chained with reverse prompting

9. Reverse Prompting

   • Before starting, agent asks 5 dynamically-generated clarifying questions
   • Surfaces non-obvious preferences, assumptions, constraints
   • Improves one-shot success rate
   • Example: “What’s the primary goal? Static or framework? Linear vs. other aesthetic? Generate copy or placeholders?”

Multi-Agent Chrome Automation (Demo)

• Each agent gets own Chrome instance + workspace + autonomy level
• Orchestrator (Claude) spawns parallel sub-agents for same task
• Example: 4 agents simultaneously scraping rental sites (Craigslist, Facebook Marketplace, PadMapper, Zillow)
• Centralized chat file for coordination; orchestrator checks every 30 seconds
• Speedup: 10 forms in ~120 seconds vs. 2-3 minutes per form serially

Context Management

Context Window Composition (in order of injection):

1. System prompt
2. claude.md/agents.md/gemini.md (user-wide rules)
3. Local project .md (project-specific rules)
4. memory.md (preferences)
5. Skills (YAML frontmatter loaded; full content on-demand)
6. Tools (MCP servers, APIs)
7. Conversation history
8. Active file contents

Quality Degradation: Model quality decreases as token count increases (e.g., 100% at 10K tokens → 40% at 199K tokens). Long context windows cause performance drop.

Compaction/Auto-Compression: When reaching ~80% of limit, models compress early context to fit new tokens—loses tool outputs, some details.

Iceberg Technique (strategic context loading):

• Above water (always loaded): Memory, learned rules, current task context, active files
• Below water (on-demand): Full codebase, file contents (via read), specific segments (via grep/glob), web data (via search)
• Use tools to selectively load only what’s needed—don’t dump entire workspace

Cost Optimization

60-30-10 Rule (Yerkes-Dodson curve for cost vs. quality):

• 60% of tokens on simplest tasks (Haiku/Flash models) → $1/M tokens
• 30% on medium-complexity (Sonnet/mid-tier) → $3/M tokens
• 10% on high-level routing/decisions (Opus/GPT-5.4) → $5/M tokens

Example cost savings:

• 100% Opus: 100M tokens × 5/M = 500
• 60-30-10 split: (60M × 1) + (30M × 3) + (10M × 5) = 200 (60% savings)

Lead scraping example:

• Scraping: Haiku at $0.001/lead
• Enrichment: Sonnet at $0.008/lead
• Outreach: Sonnet at $0.005/lead
• QA review: Opus at $0.015/lead
• Total: ~0.029/lead vs. 0.12/lead all-Opus

Batch APIs: Submit bulk requests for 50% discount; providers run during off-peak hours (load balancing).

Notes

• Course emphasizes parallelization as key advantage of AI agents—they’re not as intelligent as humans individually, but running multiple instances simultaneously achieves better results through speed and statistical coverage
• Many techniques covered not commonly discussed on YouTube (as of recording date)
• The course files, skills, and examples are linked in video description
• Nick runs a business doing $4M/year using AI agents; teaches 2,000+ people