Anton Vasilev

About the program

A hands-on live workshop for engineers working with AI agents, focused on the harness - everything wrapped around the model: the system prompt, tools, the agent loop, context management, rules, skills and hooks. The harness is what turns a model into a working agent: the same model inside different harnesses scores more than 20 percentage points apart on benchmarks. The workshop breaks down what happens under the hood of Claude Code, Codex and other agents, and where you can step into that process. Day one is the harness from the inside. Participants built an agent loop in TypeScript from scratch without frameworks: the agent searches Hacker News discussions, reads the comments and saves a digest. Step by step they added tools, message history, request tracing, prompt injection protection by marking external data, loading of AGENTS.md and rules, a skills catalog and a hook requiring approval before a file is written. Day two is the harness as a variable. How Claude Code and Codex load root and nested AGENTS.md files, how hooks block actions and hide secrets from the model, how much context MCP tools take up and how skills load progressively. A walkthrough of Pi, DeepSeek Harness and Hermes Agent, a comparison of six harnesses and how to choose one for a task. In practice, an agent on the Flue framework connected to Telegram: session memory, notes, tool argument validation, an owner-only tool and a logging hook. Outcome: an understanding of what any agent is made of and how to influence it at the prompt, context and harness levels, plus a working agent loop and a framework-based agent that can be adapted to one's own tasks. Total volume: about 6 hours of material and 2 hours of live questions and answers. Participants received access to the recording, code repositories broken down by 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 - The Harness from the Inside

   • What a harness is: system prompt, tools, agent loop and context
   • Same model, different harnesses: why benchmark results diverge
   • Pi as a minimal harness: four universal tools and bash as the interface
   • Stateless requests, context growth and caching
   • Prompt, context and harness: three levels of control over an agent
   • Practice: a minimal agent loop in TypeScript without frameworks
   • Tools, message history and protection against prompt injection
   • AGENTS.md, rules and skills in the agent's context
   • A before-tool hook and action approval
   • Q&A: participants' questions

2. 02Day 2 - The Harness as a Variable

   • How Claude Code and Codex load AGENTS.md, nested instructions and rules
   • Hooks in coding assistants: blocking actions and protecting secrets
   • MCP and skills in context: tool search and progressive disclosure
   • DeepSeek Harness: plugin architecture and trajectory tracing
   • Pi: extensions and harness events
   • Practice: an agent on the Flue framework with memory, tools and hooks, connected to Telegram
   • Hermes Agent: an autonomous agent with its own memory and skills
   • Comparing six harnesses: which one to choose for which task
   • Meta-Harness: optimizing a harness with evals and traces
   • Q&A: participants' questions