The Moving Frontier — On AI, business, and a changing world. By Dhruv Dhingra.

An AI Agent Needs a Job Description

By Dhruv Dhingra


THE MOVING FRONTIER · JANUARY 2026 RETROSPECTIVE

Autonomy is not a feature checkbox

Ask a team what its agent does and you may hear a list of capabilities: it searches, reasons, summarizes, writes, and calls tools. Ask what it is responsible for and the answer is often less clear.

That distinction matters. An agent without a defined responsibility is not a product strategy; it is a collection of potentially impressive behaviors. In enterprise software, the valuable question is not how autonomous an agent can be, but which work it can own reliably.

I think a productive starting point is to write an agent’s job description before its system prompt.

Five things every agent should know

First, the objective: what outcome counts as completion? ‘Help with research’ is not precise enough. ‘Produce a reviewable brief answering these questions with citations to approved sources’ is closer.

Second, the environment: which systems can the agent read, what context is authoritative, and how should it handle conflicting information?

Third, the authority: which actions can it take without asking, which require approval, and which are explicitly out of scope?

Fourth, the handoff: when should it stop and ask a person for missing information, make uncertainty visible, or return control?

Fifth, the standard of performance: how will we know the result is good enough, and how do we learn from failures?

Design for the exceptions

Imagine an agent handling supplier onboarding. In the ordinary case, the required documents arrive, the fields match, and the process moves forward. The demo works.

Now introduce a supplier whose legal name differs across systems, a missing beneficial-owner declaration, and a conflicting bank account. The agent’s value depends less on whether it can fill fields than on whether it recognizes that it has crossed the boundary of its role.

A good job description says what evidence it needs, which discrepancies it may resolve, and which exceptions must be escalated. It makes the failure mode a design input rather than a surprise after deployment.

Measure ownership, not motion

Counting tool calls or autonomous steps tells us little about product value. An agent that performs 40 actions and leaves a person untangling the outcome is not necessarily better than one that performs six and produces a defensible result.

Measure completion quality, rework, exceptions handled correctly, and the time a human must spend supervising. Track the cost of an incorrect action, not just its frequency.

An agent becomes useful when the organization can assign it responsibility within explicit bounds. That is closer to designing a role than launching a feature.

Editorial note: This is a retrospective essay written in November 2026. January 2026 describes the period explored, not its original publication date.


← All essays