Module 5.1 · Topic 4
The Human-Agent Collaboration Model
Bottom Line Up Front Agents are most effective when humans remain accountable and maintain oversight. Define which tasks agents own end-to-end, which require human checkpoints, and what conditions trigger escalation.…
4.1 Delegation Patterns: What to Hand Off and What to Keep
Not all work should be delegated to agents. Your judgment determines delegation decisions. A few clear criteria separate delegatable from non-delegatable work.
| Task Dimension | Delegatable to Agent | Retain for Human |
|---|---|---|
| Rule-Based vs. Judgment-Based | Clear, consistent rules: "Export Q3 data from dashboard, format as CSV, name with timestamp" — delegate. | Requires judgment: "Decide whether this candidate is a cultural fit" — human only. |
| Reversible vs. Irreversible | Easy to undo: agent books a meeting but double-books — you cancel one, no harm done. | Hard to undo: agent submits a legal filing or publishes a public announcement. |
| Low-Stakes vs. High-Stakes | Modest impact if wrong: agent misspells a vendor name in a draft — you fix it before sending. | High consequences for error: agent grants access to sensitive systems, or makes financial commitments. |
| High Confidence vs. Low Confidence | Proven agent performance on this task type with success rates above your risk threshold. | Novel or poorly understood scenarios where agent performance is uncertain. |
4.2 Supervision and Oversight Requirements
Effective supervision prevents agent failures from cascading into disasters. Design checkpoints where you review progress before harm can occur.
Checkpoint Types
- Pre-Execution Review: You review the agent's plan before it acts. "The agent proposes to archive 500 emails matching criteria X — do you approve?" This is most effective for high-stakes operations but slows down the agent.
- Output Verification: The agent acts, then you review results before they take effect. "Agent extracted 50 customer records — verify the sample and approve before importing into CRM." This is practical for most delegations.
- Exception Handling: The agent acts freely but alerts you if it encounters exceptions. "Agent processed 1,000 invoices; 47 had parsing errors and were set aside for manual review." You handle the exceptions; routine work completes unsupervised.
Supervision Pitfalls
Alert Fatigue: Too many alerts and you stop reading them. Reserve alerts for genuinely exceptional conditions, not routine notifications.
Rubber-Stamp Approval: If you approve agent work without actually reviewing it, supervision is theater. Spot-check samples, not every result. If results prove unreliable, increase review frequency.
Delayed Escalation: If you don't review outputs promptly, errors cascade. Schedule regular review windows (daily, weekly) depending on task frequency and impact.
4.3 Trust Boundaries and Escalation Triggers
An agent should not proceed blindly. Define explicit conditions that cause it to stop and ask for help. Escalation triggers prevent the agent from exceeding its authority.
- Confidence Threshold: If the agent's confidence in its next action drops below a threshold (e.g., 70%), escalate. This prevents the agent from fumbling when uncertainty rises.
- Financial Threshold: If the operation would move more than X dollars, cost more than Y, or affect Z customer records, escalate for approval.
- Retry Limit: If the agent has tried the same action three times and failed, escalate rather than loop endlessly.
- Permission Denial: If the agent is denied access (permission error, authentication failure), escalate immediately. It may need elevated credentials, or the operation may be unauthorized.
- Data Anomaly: If the agent receives data that violates expected patterns — a customer balance that's negative, a file size that's orders of magnitude larger than expected — escalate for investigation.
- Timeout: If the agent has been working on a single task for longer than expected, escalate. It may be genuinely blocked or stuck in a loop.
4.4 The Evolving Role of the Human Operator
Agent deployment changes your job. You transition from doing the work to managing the agent doing the work. This is a profound shift that requires new disciplines.
From Executor to Delegator. You no longer execute every task yourself. Instead, you decompose goals into agent-friendly subtasks, define success criteria, and hand off work. This requires writing clear specifications — a skill distinct from doing the work itself.
Accountability Remains Yours. The agent acts on your behalf. You bear responsibility for the agent's mistakes. This means rigorous oversight, not blind trust. Review work before it reaches stakeholders. Maintain audit trails showing what the agent did and when. If something goes wrong, you explain why it happened and what you'll do differently.
Judgment and Risk Management Intensify. You focus on higher-value work: evaluating whether agent outputs are correct, deciding which tasks warrant agent delegation, and managing the risk that agents introduce. You become a quality controller and risk manager, not a task executor.
Continuous Improvement Becomes Standard. After each major agent run, you review what worked and what didn't. Did the agent succeed? Did it encounter unexpected obstacles? Should you adjust the specification, train the agent differently, or choose a different approach entirely? Feedback loops turn agent failures into lessons.