Decoding AI’s Open-Source Course Maps Three Ways to Run Agents

When harness design dictates quality, executing the agent loop becomes an architectural choice rather than a mere deployment detail. Paul Iusztin's open-source course, Building a Coding Agent From Scratch, builds a Python agent named Decode through Decoding AI. The system separates three distinct run modes, each featuring a unique latency profile and corresponding inference provider requirements.

At the center lies a headless harness devoid of its own interface, executing a shared agent loop: the LLM selects an action, a tool executes, and observations feed back into the context window. The agent itself remains lightweight; in Decode, it is roughly a 20-line Pydantic AI definition, while Claude Code's core loop spans approximately 150 lines. All other components—including memory, skills, sandboxes, permissions, and LSP feedback—reside within the harness.

Stock photo for illustration only, not from the actual event

Interfaces plug directly into this core engine, giving rise to three distinct operational modes:

• Terminal UI: Tied to a single live session in memory within the same process, streaming events via async generators.
• Remote Mode: Keeps the harness headless, running on servers via Kitaru, ZenML on GCP, and executing on Modal.
• Background/Async Mode: A live session hands off work to a job queue and returns immediately, allowing background workflows to process LLM calls asynchronously.

Addressing the challenge of steering in live terminal sessions, Decode implements a steering queue paired with a priority gate. Inputs arriving while a tool call is in flight are buffered and injected strictly at safe boundaries, defined as MODEL_REQUEST and WOULD_STOP, supporting standard Enter, Alt+Enter, and Esc abort commands.