Getting started with AI agents w/ Oleg Podsechin & Toughbyte
Getting more out of AI agents — Oleg Podsechin at Maria 01
Most companies are barely scratching the surface of what AI can do. This session cuts through the hype and shows what actually works.
Oleg Podsechin (Toughbyte, Ghostlyn) breaks down why LLMs are just next-token predictors — powerful, but not magic. The real gains come from the harness: memory, tools, agentic loops, and careful task selection. He walks through the METR horizon framework, why tool choice matters more than model choice, and how retrieval-augmented generation and fine-tuning solve very different problems.
He also covers the security angle: why you should never hand a fully autonomous agent the keys to production, how "sandbox escape" became the top AI risk in 2024, and what the lethal trifecta of AI security looks like in practice.
The back half turns practical. Oleg explains why foundation models are becoming prohibitively expensive, how regulation is locking down model access, and why rolling your own agent stack is more friction than most teams realize — then shows how Ghostlyn's team-first, model-agnostic approach fixes all three.
0:00 Intro & agenda
2:50 Meet the speakers — Toughbyte & A foundersfriend
7:00 Hype vs reality: where AI value actually comes from
8:40 LLM basics: next-token predictors that hallucinate
12:30 Common misconceptions about LLMs
14:30 What is an AI agent? Model, harness, tools, context
17:00 The agentic loop
19:30 Why agents are taking off now
22:30 Choosing a model: open weight vs proprietary
24:30 Harnesses: build your own vs use existing
27:00 Tools and function calling
30:00 Context windows and compaction
33:30 Sub-agents and offloading context
38:00 Memory, RAG and embeddings
40:00 Agent security & prompt injection
41:30 Agent skills and progressive disclosure
44:30 One agent or many?
47:30 Inside Ghostlyn: prompt, tools, memory, model
51:30 Connecting to external services (Meta, Gmail)
55:00 Do you actually need an agent? Workflows vs agents
57:30 Answer agents vs action agents & human-in-the-loop
1:00:00 Picking what to automate
1:03:00 Common pitfalls and skill degradation
1:06:30 Skills demo: Hacker News digest & support skill
1:11:00 What's next: faster inference, dynamic MCP
Speakers;
Oleg Podsechin
Nicolas Dolenc
Links:
Ghostlyn - ghostlyn.com
Toughbyte — toughbyte.com
A Founder's Friend — afoundersfriend.com
#AIAgents #LLM #Ghostlyn #Toughbyte #Maria01 #AIEngineering