When a coding agent starts losing track of a task, should you compact the context, start a fresh session, create a subagent, change the instructions, or give it another tool?

Most developers learn answers to questions like these as rules of thumb. But those rules become much easier to apply and debug once you understand the mechanics underneath the agent.

This session builds a systems-level mental model of modern coding-agent harnesses using tools such as Claude Code and Pi as examples. We will start by separating the model from the harness around it, then follow how a task actually moves through the system: how the harness constructs context, how the model generates the next action, how structured outputs become tool calls, how tool results return to context, and how this process creates an agent loop.

From there, we will use that model to explain practical behaviors engineers encounter every day: why context is working memory rather than storage, what compaction actually changes, why persistent instruction files work, when subagents help through context isolation, where MCP and tools fit into the architecture, and why tests, compilers, linters, and other deterministic systems are critical for verifying probabilistic agents.

The goal is to give software engineers a durable mental model they can use to understand, debug, and engineer effectively with any coding-agent harness.