If a manager is accountable for an outcome, they need enough understanding to question how the people and agents producing it actually work.
An emerging responsibility
We use agent manager to describe a management responsibility, not a universally established job title. An AI agent can use tools and data to pursue a defined task. Several agents may exchange information or hand work to one another. A manager needs to understand the whole workflow, its boundaries, and who owns its consequences.
Slide decks, meetings, and status summaries still have a place. They are insufficient on their own when work is being executed by systems between meetings. The manager needs a view of the decisions, exceptions, and operating evidence underneath the summary.
Delegation becomes more explicit
A human colleague brings context and can negotiate an unclear instruction. An agent’s operating scope needs explicit definition: the objective, approved data, available tools, spending limits, acceptance criteria, and actions that need approval.
Delegating to a network of agents also creates dependencies. One incorrect classification can shape a retrieval step, a recommendation, and a downstream action. Someone must own the end-to-end result and the rules for handing work across those boundaries.
Learn to read the evidence
Managers do not need to become full-time engineers to ask good operational questions. They do need enough fluency to read a run history, distinguish an error from a business exception, inspect a data source, and understand what a test does and does not establish.
A log records what a system reports happened. An evaluation checks performance against defined examples and criteria. A business measure tells you whether the workflow improved the intended outcome. Each answers a different question; none is a complete picture on its own. Engineering partners remain essential for instrumentation, reliability, security, and incident response.
Around-the-clock work needs around-the-clock boundaries
An agent workflow can be configured to run outside business hours. That does not make every task suitable for unattended operation or guarantee availability. Set rules for queueing, approvals, alerts, retries, and stopping. Agree who responds when an exception exceeds the permitted scope.
For example, an overnight process might gather evidence and draft a service recommendation, then wait for an authorized person before making an external commitment. If the evidence is incomplete, the correct outcome may be to stop and flag the case. More autonomous activity is not automatically better performance.
Change the scorecard
Count accepted outcomes, not just generated outputs. Track quality, error severity, rework, human review time, escalation quality, cost per accepted outcome, and customer or employee effects. Compare with a baseline for the same type of work.
Return on investment should account for implementation, integration, evaluation, supervision, and operating costs alongside realized benefits. Released capacity becomes value only when the organization can put it to use. A fast draft that shifts hidden work to another team can look productive while making the overall process worse.
Keep human leadership central
Managers must explain changing responsibilities, give people time to learn, and invite challenges to a system’s decisions. They need to recognize when a team is absorbing more exceptions or stress even while the headline throughput rises.
Thinkerland sees agent leadership as a combination of people management, process ownership, operational literacy, and responsible experimentation. The aim is a manager who can define useful work, inspect the evidence, intervene competently, and develop the people working alongside the systems.
Could the accountable manager explain what happened in the last failed agent run—and what should happen next?
Research and further reading
Sources that inform the conversation. The recommendations above express Thinkerland’s point of view.
- NIST AI RMF Playbook: Measure
Operational context: voluntary practices for monitoring, error investigation, and human oversight; not a certification.
- Microsoft 2026 Work Trend Index
Industry perspective on human agency, workflow redesign, and agent oversight. Vendor research, not a forecast for every organization.
Put the idea to work.
Examine a sample agent workflow, the manager’s decisions, and a more useful performance scorecard.
Explore agent leadership