Agentic systems don’t replace humans—they change where humans show up. The key design question: at what points does a human need to be in the loop, and what do they do when they’re there?

Three patterns structure that design: Delegation, Supervision, and Handoff. Getting them right is the difference between an agent that feels like a helpful colleague and one that feels like a black box that sometimes works.

Delegation

What it is: The human assigns a task to the agent with clear boundaries. The agent executes; the human gets the result.

Design considerations:

  • Clarity of scope: The agent needs to know what’s in and out. Vague delegation leads to scope creep or irrelevant output.
  • Reversibility: Can the human undo or override? Delegation without recourse creates anxiety.
  • Feedback: Does the human see progress, or only the final result? Long-running tasks need visibility.

Knowledge for humans: What to delegate (well-defined, lower-stakes tasks) vs. what to keep (ambiguous, high-stakes decisions). When to add constraints vs. when to leave room for agent judgment.

Supervision

What it is: The human observes agent activity and intervenes when something goes wrong or when approval is needed. The agent runs; the human watches and occasionally steps in.

Design considerations:

  • Intervention points: Where does the agent pause for human input? Too few, and the human can’t catch errors. Too many, and the human becomes a bottleneck.
  • Signal quality: What does the human see? Raw logs vs. summarized status vs. highlighted anomalies. Good supervision requires good visibility.
  • Override protocols: How does the human interrupt? Kill the agent? Redirect it? Feed it new instructions? The protocol matters for trust.

Knowledge for humans: What to watch for. When to let the agent continue vs. when to intervene. How to intervene effectively (what input actually helps).

Handoff

What it is: The agent completes a phase and passes work to a human or another agent. The handoff includes context, artifacts, and recommendations. The recipient picks up and continues.

Design considerations:

  • Context transfer: What does the recipient need to know? The agent’s reasoning? Its confidence? What it tried and failed? Handoffs fail when context is lost.
  • Artifact quality: What’s being handed off—code, a doc, a plan? Is it in a usable state, or does the recipient need to fix it first?
  • Ownership: Who owns the outcome after handoff? Clarity prevents “the agent broke it” vs. “I broke it” finger-pointing.

Knowledge for humans: How to receive a handoff—what to validate, what to assume is done, what to re-do. How to hand back to the agent when the human’s part is done.

Combining the Patterns

Real agentic systems layer these patterns. A development agent might delegate a sub-task to a code-generation sub-agent, supervise the overall plan with human checkpoints, and hand off to a human for code review before merge. The design challenge is choosing the right pattern at each boundary—and teaching humans the knowledge to operate within it.