Delegate vs Collaborate as the Meta-Skill of Knowledge Work
Katie Parrott’s Codex guide collapses every agent task into a binary choice. Delegate when the work is predictable, repeatable, and low-risk; with clear, well-specified instructions the agent executes autonomously and brings back finished work for review. Collaborate when the work is judgment-heavy, exploratory, or iterative; you stay in the loop and steer alongside the model toward an outcome that matches your vision.
The strong claim is in the next sentence: “Deciding which mode fits which needs is the meta-skill of modern knowledge work.” The bottleneck is no longer doing the task or even prompting well. It is sorting a stream of incoming tasks into the right bucket, because miscategorising punishes you in both directions. Delegate a judgment-heavy task and the agent fills the gaps with confident wrong guesses. Collaborate on a repeatable one and you become the agent’s bottleneck for work that should have run while you slept.
The taxonomy sits above task-by-task technique. It is the sentence you say to yourself before opening the agent.
Related
- Three Modes of Working With AI — Boris Cherny’s sibling taxonomy (vibe coding / pairing / by hand) splits by output type and human craft; Parrott’s splits by autonomy level and judgment density
- Your automation isn’t set and forget, every agent needs a human on both ends — the “human sandwich” frame; the delegate mode still requires framing and judging, just compressed to the two ends
- A Plan Document Can’t Hold Your Product Vision — the cost of miscategorising as delegate when the work is actually collaborate: no spec can carry judgment that has not been formed yet
- Loops Make Sense Only With a Fixed Feedback Signal — a sharper criterion for the delegate side: autonomous loops only work when “well-specified” includes a machine-checkable score