▸ EX Epic Academy · Module 01 · Free Core Curriculum

Agentic AI Fundamentals.

The shift from asking an AI a question and getting an answer, to giving an AI a goal and having it work autonomously until the goal is met. Everyone needs to code. Period.

Free · Core8 ChaptersReading · 14 min
▸ Chapter 01 · The Definition

What Agentic AI actually is.

Agentic AI is the shift from asking an AI a question and getting an answer, to giving an AI a goal and having it work autonomously until the goal is met. The AI doesn't just respond it plans, uses tools, writes files, runs commands, spawns sub-agents, and iterates until the job is done.

This is not a better chatbot. This is a digital worker that ships code, writes research, manages infrastructure, and coordinates with other agents while you sleep.

01

Tool use

The AI reaches into your world terminal, browser, filesystem, APIs, databases. It doesn't describe what you should do. It does it.

02

Autonomous loops

The AI thinks, acts, observes the result, and decides the next step. This repeats until the goal is met or a limit is hit. One prompt can trigger 5, 10, 50 tool calls with no human in the loop.

03

Persistent context

The AI remembers across sessions who you are, what you're building, your preferences, your conventions. Every interaction builds on the last.

▸ Chapter 02 · The Current Systems

Three architectures, one pattern.

All modern agentic systems follow the same core loop: LLM thinks → calls tools → observes results → decides next step. But they differ in scope, platform, and philosophy.

Hermes Agent

Nous Research The Platform

The most ambitious of the three. Not just a coding agent a full agent operating system.

Provider-agnostic

Swap models mid-session. DeepSeek, Claude, GPT, Grok, Gemini, local models all work. No lock-in.

Self-improving via skills

When Hermes solves a hard problem, it saves the approach as a reusable skill. Over weeks, it accumulates hundreds of skills specific to your work.

Multi-platform gateway

The same agent runs on Telegram, Discord, Slack, WhatsApp, iMessage, SMS, email with full tool access. Text it from your phone and it builds software on your machine.

Persistent memory

Remembers who you are, what you prefer, your environment. Survives restarts and model changes.

Profiles

Multiple independent Hermes instances with isolated configs, sessions, skills, and memory. One agent for work, one for personal, one for a project.

Cron

Scheduled autonomous jobs. Morning briefing, daily research ingest, weekly self-learning, memory compaction all without you.

Kanban swarm

Six specialist profiles (researcher, analyst, writer, reviewer, backend-eng, ops) that dispatch tasks from a shared work queue.

Delegation

Spawn sub-agents for parallel work. Three agents working three files simultaneously, then reporting back.

The system we run here at EX Venture. The backbone. But not the only tool.

Claude Code

Anthropic The Specialist

The most polished coding agent on the market. Narrower than Hermes code only but exceptional within that domain.

Two modes

Print mode (-p) for one-shot tasks with structured JSON output. Interactive mode for multi-turn REPL sessions with TUI.

Deep code understanding

Reads entire codebases, understands architecture, refactors across dozens of files in a single session.

Sub-agents

Spawn specialist sub-agents (security reviewer, database expert, test writer) and orchestrate them from a parent session.

Hooks

Automation triggers on tool use auto-format on write, block dangerous commands, log completions.

MCP integration

Connect to external tools databases, APIs, GitHub through the Model Context Protocol.

Worktree isolation

Run multiple Claude instances on the same repo in isolated git worktrees. No conflicts.

Locked to Anthropic models. Brilliant, but you can't swap to DeepSeek for cheap inference or GPT for a second opinion.

Codex

OpenAI The Hacker

OpenAI's entry. The youngest and most raw of the three, but with one thing the others don't: true sandboxing.

Exec mode

One-shot task execution. Give it a prompt, it builds the thing, exits.

Sandbox modes

Full-auto sandboxes file changes. Yolo mode skips all approvals. Danger-full-access runs with no sandbox at all.

Git-native

Refuses to run outside a git repo. Every change is trackable, revertible.

Lightweight

Fewer features than Hermes or Claude Code, but faster to launch and simpler to reason about.

Still maturing. Fewer hooks, no profiles, no memory comparable to Hermes. But for quick "build this" tasks, it's fast and clean.

▸ Chapter 03 · Multi-Model · Code-First

Why picking one model is the biggest mistake.

Different models have different strengths. A code-first multi-model approach means the orchestrator picks the right model per task, runs parallel agents on different models simultaneously, and falls back automatically when a provider is overloaded.

Claude Sonnet 4

Complex reasoning, architecture decisions, code review nuance.

DeepSeek V4 Flash

Fastest, cheapest, good enough for 80% of tasks. Our default.

GPT-4.5 / O3

Long-horizon planning, creative problem framing.

Grok

Fastest at search-grounded tasks, X data access.

Gemini

Massive context windows, multimodal reasoning.

01 · Orchestrator picks the model

Hermes picks the right model per task. Flash for routine work, Pro for complex reasoning, Claude for architecture, Grok for research.

02 · Parallel agents, different models

Three agents on three tasks, each using the model best suited to their work simultaneously.

03 · Model fallback

If one provider is overloaded, the system falls back to another automatically. No single point of failure.

04 · Cost optimization

Flash costs a fraction of Sonnet. Use Flash for 80% of turns, Pro/Claude for the 20% that need depth.

The architecture: Hermes is the platform layer the operating system. Claude Code and Codex are specialist tools that Hermes orchestrates. One platform, many models, tool-agnostic execution. This is what we run.

▸ Chapter 04 · Under the Hood

How agentic systems actually work.

Strip away the marketing and every agentic system has the same skeleton eight load-bearing pieces.

01

System prompt

The agent's identity, rules, constraints. SOUL.md in Hermes, CLAUDE.md in Claude Code. Tells the agent who it is and how to behave.

02

Tool layer

The interface between the agent's thinking and the real world. Terminal commands, file reads/writes, browser navigation, API calls, sub-agent spawning.

03

Agent loop

The core cycle: input → think → decide on tool calls → execute → observe → think again. Loops until the goal is met or max_turns is hit (default 90 in Hermes).

04

Memory

Persistent facts that survive sessions. Preferences, environment details, corrections. ~2,200 character storage, FTS5 indexed, survives restarts.

05

Skills

Procedural knowledge reusable approaches for recurring tasks. 100+ skills covering research, coding, patent drafting, novel writing, hardware sourcing, grant applications. They accumulate.

06

Cron

Scheduled autonomous execution. No human trigger. Research ingestion at midnight. Memory compaction on Sunday. Morning briefing at 7 AM.

07

Profiles

Complete agent isolation. Different config, skills, memory, model. Work agent and personal agent on the same machine, zero cross-contamination.

08

Delegation

Sub-agents with isolated contexts. Child inherits parent's model, tools constraint, and goal then works independently and reports back.

▸ Chapter 05 · The Future

Where this is going.

We are at the beginning. Here's what's coming within the next 18 months.

01 · Agent-native operating systems

Hermes already functions as a proto-OS agent loop, process management, filesystem, scheduler, multi-user gateway. The line between "AI tool" and "operating system" blurs within 18 months.

02 · Swarm intelligence

Six Kanban profiles today. Fifty tomorrow. Agents that specialize, negotiate, and coordinate without human routing triggered by one sentence from you.

03 · Continuous learning

Skills that self-update based on failures. Memory that consolidates and prunes automatically. Agents that wake up smarter than they were when you went to bed.

04 · Hardware integration

Agents that control physical systems 3D printers, CNC machines, sensor arrays, robotics platforms. The agent loop extends into the physical world.

05 · Economic agents

Agents with budget constraints that make cost/benefit decisions. "Build this feature, but don't spend more than $5 in API costs."

06 · Trust and verification

Sub-agents report success, but did they actually succeed? Cryptographic verification chains every file write traced, every claim cross-checked, every deliverable validated before you see it.

▸ Chapter 06 · What This Means for EX Epic Academy

You are not learning a tool. You are learning to work with a digital workforce.

Module 1 is about understanding the paradigm. Not the buttons the mental model.

The skill is not "prompt engineering." The skill is delegation, orchestration, trust calibration, and system design.

A single human with a well-tuned agent fleet can outproduce a ten-person team. This is already true today. The gap will only widen.

The Rest of EX Epic Academy

Teaches you how to build that fleet architectures, memory systems, skill libraries, multi-agent orchestration, and the economic framing to run agents at scale.

Explore the Academy →

EX Venture Group and its affiliated service companies are incorporated and operate in the United States and Europe only. They do not operate in Indonesia and do not provide services into or from Indonesia. All Bali-based operations, including EX Epic Academy, are solely managed and operated by EPIC Solutions PT.

© 2026 EX Epic Academy · Operated by Epic Solutions Global, Bali, Indonesia

▸ UU PDP No. 27/2022 · UU ITE · Cookie consent required

▸ Cookies & Privacy

We use essential cookies to operate this site and analytics cookies to improve it. In line with Indonesia’s UU PDP (Personal Data Protection Law, UU No. 27 of 2022), your consent is required for non-essential cookies. See our Privacy Policy.