Orchestrating Coding Agents

What you’ll learn and how you can apply it

• Diagnose which limit you’re actually hitting (context, serial execution, model cost, or a single point of failure) and pick the pattern that addresses it
• Build a handoff loop with a scratchpad file, so work survives across fresh context windows without silent summarization
• Run a generator-critic review with a defined pass condition and a hard pass limit, and filter the findings before acting
• Set up parallel agents on one repo with Git worktrees, so each one ends up as its own reviewable pull request

Course description

Orchestrating agents involves the practice of running more than one coding agent on the same work and deciding which agent does what, in what order, and how the results come back together. Engineers reach for it when a single agent stops being enough, and the signs are consistent: the context window fills up and the agent drifts from its own plan, independent tasks run one at a time, one expensive model handles both the architecture decisions and the mechanical renames, and an assumption made wrongly early on has nothing in the system to catch it. Orchestrating multiple agents addresses those limits and introduces new ones, including merge conflicts between agents editing the same files, errors that compound across handoffs, and token costs that grow faster than the agent count.

This course with Alfonso Graziano covers the main orchestration patterns in order of complexity: a scratchpad file that carries state into a fresh context window; a second agent that reviews the first one’s work under defined stop rules; several agents running in parallel on one repository with context, worktree, or machine isolation; subagents in their four common arrangements; and orchestration written as a script so a repeated flow runs the same way every time. Each pattern is demonstrated live on a real repository. By the end of the course, you’ll be able to diagnose which limit you’re hitting, choose the lightest pattern that addresses it, and set up parallel agents so each stream of work ends as its own reviewable pull request.

The course is discussion-based, with live demos run by Alfonso, so no setup is necessary. The scripts and prompts used in the demos are available as downloads during the session, so you can run them afterwards on your own repo.