AI Workshop
Episode 3 of 10Aired

How an agent actually works

The «think → call a tool → read the result» loop, the context window, and why an agent suddenly «forgets».

Wednesday 21 October at 19:00 Moscow time· 90 min

Aired

The recording is being prepared — it usually appears within two days. Sign up for the next streams so you don't miss them.

What we'll show

  1. The agent loop step by step on a live task: reasoning, tool call, result, next step
  2. The agent's tools: reading and editing files, the terminal, search, MCP
  3. The context window: what goes in, how it fills up, what happens on compaction
  4. Why a long session degrades and when to start a fresh one
  5. Reading the agent's log to see where it took a wrong turn

What you take away

A working model of the agent loop — so you predict its behaviour instead of being surprised by it.

Cheat sheet: the agent loop and context

  • Yes: An agent is a model in a loop: reasoning → tool → result → reasoning again
  • Yes: Everything the agent «knows» about the project is what's in context: files, command output, CLAUDE.md, your messages
  • Yes: Long command output eats context: ask the agent to filter (grep, head), not print everything
  • Yes: One session, one task; a new task means a new session or /clear
  • Yes: If the agent «forgot» an agreement, it isn't in context — move it into CLAUDE.md
  • Yes: Read the log: at which step did the agent decide wrong, and what information was missing

Harness in this episode

The harness is what surrounds the loop: what the agent sees going in and what checks the result coming out.

The full harness episode