Agent Orchestrator
You program multi-step autonomous processes. You spawn subagents. You direct rather than do.
- You write scripts that call Claude's API, chain tasks, and handle outputs programmatically
- You spawn subagents and program multi-step autonomous processes
- The AI executes plans; your role is specification, review, and intervention
This is the only level where the AI can cause downstream effects — file changes, API calls, published outputs — without you reviewing each step. The chain completes; the output looks plausible; something went wrong at step 3. Verification architecture must be built into the workflow design, not added afterwards. Trust is earned through iteration on small runs, not assumed from the start.
Level 4 is the full scope of current AI-assisted research practice. What matters now is not moving further but using this level well: designing robust verification architectures, documenting your workflows, and staying alert to the ways that autonomous systems introduce error modes that don't exist at lower levels.
There is a pull toward Level 4 in how AI is discussed publicly. "Agent orchestration" sounds more impressive than "I have a good CLAUDE.md." This is a distortion. The highest-leverage gains for most humanities researchers come from Level 2 (persistent context) and selective Level 3 workflows for file-heavy tasks. Level 4 is valuable in specific circumstances: large-scale corpus work, institutional tasks that are genuinely routine, researchers with programming fluency who have a concrete automation problem. It is a tool, not a destination.