Olexiy Nakhod

About the program

A hands-on live workshop for engineers working with AI agents, focused on the harness - the whole program around the model: the system prompt, tools, the agent loop, context, permissions and hooks. The model only generates a tool call request; it is the harness that reads files, runs commands, decides what may be executed and when to stop. Cost and results depend on it even with the same model. Day one is ready-made harnesses from the inside. A master class with live demos in Claude Code and Codex: what sits in the context before your first word, how AGENTS.md, rules and skills are loaded, how deny and ask permissions, the sandbox, guard hooks, the Stop hook and mods work, and how long sessions handle compaction, subagents and memory. An overview of the open harnesses Pi, Hermes Agent and DeepSeek Harness, and of research on how the harness affects cost and quality. Day two is building your own harness. Participants built a harness in TypeScript in 11 stages, from the first model request to a note-taking Telegram bot: the agent loop, tools with argument schemas, step and budget limits, failure handling and retries, conversation memory, a skill, permissions through a pre-tool hook and a stable context prefix for caching. Every step of the loop is visible live in a visual trace. Outcome: an understanding of what any coding agent is made of and which setting controls what, plus a working harness of their own and a Telegram bot built on it that can be developed further for their own tasks. Total volume: about 6 hours of material plus live questions and answers at the end of each day. Participants received access to the recording, a repository with the code for every stage, step-by-step guides for each stage and environment setup instructions.

Instruction language: Ukrainian

Learn more about the workshop →

Skills

• Artificial Intelligence (AI)
• AI Agents
• Large Language Models (LLM)
• Prompt Engineering
• TypeScript
• Node.js
• Claude Code

Curriculum

1. 01Day 1 - Ready-Made Harnesses from the Inside

   • What a harness is: the agent loop and the decisions the harness makes instead of the model
   • Same model, different harnesses: how the harness moves cost and quality
   • What the model sees before your first word: /context, deferred MCP tools and caching
   • Vendor system prompts, code search and AGENTS.md as a shared standard
   • How Claude Code and Codex load AGENTS.md, rules and skills
   • Long sessions: compaction, subagents and memory
   • Control in Claude Code: deny and ask, the sandbox, guard hooks, the Stop hook and mods
   • Open harnesses: Pi, Hermes Agent and DeepSeek Harness
   • Q&A: participants' questions

2. 02Day 2 - Building Your Own Harness

   • Setup: a web harness with a trace of every step
   • The model: the first request, tokens and cost
   • The agent loop and the first tool
   • Tools: argument schemas, call ids and invalid calls
   • Limits: step and budget caps, truncating tool output
   • Failures: errors as data, retries and truncated responses
   • A Telegram bot on the same harness
   • Conversation memory for every chat
   • A skill: a catalog in the system prompt and a readSkill tool
   • Permissions: a pre-tool hook and owner-only tools
   • Caching and AGENTS.md: a stable context prefix
   • Q&A: participants' questions