Key Takeaways
- Agentic AI systems go beyond conversational prompting by persistently holding context, capabilities, and analytical orientation across entire data sets.
- Professionals can configure multiple agents with different orientations to generate competing insights; friction between those outputs reveals questions no single analysis would surface.
- Four discovery tactics—using multiple lenses, surfacing silences, bridging levels, and stress‑testing categories—leverage this friction to uncover hidden patterns, dependencies, misallocations, and flawed classifications.
- The value of agentic output lies in treating it as a proposal that opens inquiry, not as a final answer; professionals must judge, verify, and track which patterns they pursue.
- As orchestration layers mature, agents can trigger one another autonomously, shifting the expert’s role from configuring individual bots to shaping the conditions under which the system discovers productively.
Overview: From Prompting to Directing Intelligence
Conversational generative AI lets professionals ask questions, evaluate replies, and refine prompts—a process that hinges on articulation and the ability to keep an analytical thread in mind. Agentic AI, by contrast, is “configured and directed rather than conversed with,” enabling it to hold more data and sustain analytical orientations across whole data sets without losing the thread. The authors argue that the most valuable professional output is insight—a genuinely new way of seeing a problem—rather than faster summaries. Expertise, while useful, can blind professionals to edge‑case insights; agentic systems help surface those edges by creating friction between competing interpretations, data and discourse, causal levels, and categorical realities.
Two Ways of Working With AI
When professionals treat AI as a conversation, they must constantly supply context and keep the analytical thread alive, relying on skill in “knowing what to ask, how to phrase it, and when to push back.” Agentic interaction flips this dynamic: the user configures the system, and it operates proactively. Three configuration choices define an agent’s behavior:
- Context – what the agent can access (persistent connections to databases, documents, records).
- Capabilities – what the agent can do (run analyses, query data sets, invoke tools without waiting for human input at each step).
- Orientation – what the agent pays attention to (an analytical directive that sets purpose and trajectory).
Because orientation shapes how an agent encounters data, the same context and capabilities, given different orientations, will surface different patterns. A single professional can therefore direct multiple agents against the same data set and obtain genuinely different discoveries—not by asking different questions but by designing different systems. An orchestration layer often compares and synthesizes the agents’ outputs, allowing the professional to focus on evaluating what the system reveals rather than maintaining the analysis.
Four Approaches to Discovery With Agentic AI
Discovery emerges from friction—putting things in contact that are normally kept apart. The article outlines four specific moves that create such friction, each realizable through prompting but deepened by agentic architecture.
Use Multiple Lenses – Applying competing frameworks simultaneously highlights contradictions that no single analysis would find. In the metal‑parts example, a Porter agent, a VRIO agent, a Rumelt agent, and a Martin agent each examined the same data set through a different strategic lens. The orchestration agent surfaced the tension that the firm’s genuine advantage in aerospace was diluted by spreading investment evenly across all segments, prompting the question: “Should the company narrow its scope to the one segment where its genuine competitive advantage meets structurally attractive conditions?”
Surface Silences – Comparing what data contains with what the organization discusses reveals unspoken dependencies. In the professional‑services case, the agent noted that roughly a third of interview transcripts referenced the practice leader’s judgment, relationships, or standards, yet strategic plans never named him. The resulting insight shifted the growth question from “Can this practice grow?” to “Can it grow without first solving the problem nobody has put on the table?”
Bridge Levels – Tracing causes across organizational levels exposes misalignments between symptoms and their roots. For the diversified industrial corporation, the agent linked declining ROIC to a corporate capital‑allocation formula that overinvested in the fast‑growing, low‑margin Division A while starving the high‑margin Division B. The insight showed that the “ROIC decline the CEO has been attributing to the market is actually being produced by a corporate policy operating exactly as designed.”
Stress‑Test Categories – Mapping formal classifications against operational reality uncovers hidden causes and mislabelled rejections. The ready‑mix concrete analyst found that many “late‑delivery” rejections were actually caused by a weather‑sensitive batch plant hopper, and that a recurring “site not ready” driver comment had no category at all. The agent’s exhaustive pattern analysis revealed that “the categories are both imprecise and actively misleading,” redirecting investigation to the plant and to coordination gaps with subcontractors.
How to Direct Intelligence Skillfully
To reap the full benefit of agentic systems, professionals must adopt a discovery mindset rather than an answer‑seeking one. The authors offer several guiding principles:
- Configure for discovery, not answers. Rather than telling an agent what to find, set it up to hold competing perspectives in tension. An agent configured merely to confirm a presumed advantage will miss the insight that the advantage is real but diluted by incoherent investment.
- Treat the unexpected as signal, not error. When agents produce outputs that diverge from expectations, resist the urge to dismiss them as mistakes. Those surprises often point to hidden assumptions or structural truths, as with the “site not ready” pattern that had existed in driver comments for years.
- Evaluate proposals, not conclusions. Agentic patterns are proposals that open inquiry, not final findings. Verification—checking GPS timestamps, driver logs, or alternative explanations—turns a proposal into actionable knowledge.
- Track what you advance, and what you reject. Over time, logging which proposals are pursued and which are set aside builds a record of analytical instincts, revealing contexts where judgment either fuels or forecloses inquiry.
The authors stress that “Patterns surfaced by an agentic system are proposals, not findings. They open inquiry rather than close it.” This shift from problem‑solving to problem‑setting is central to directing intelligence effectively.
The Evolving Role of Orchestration
Current agentic designs already rely on an orchestration layer that compares and synthesizes specialist agent outputs. As these systems mature, orchestration can become more autonomous: a silence‑surfacing agent might trigger a lens‑multiplying agent to dig deeper, or a bridging agent could hand a pattern to a category‑testing agent without human prompting. In such a future, the professional’s role shifts from designing individual configurations to shaping the conditions under which agentic systems discover productively. The underlying skill—framing problems so that new things become visible and holding contradictions open long enough to learn from them—remains unchanged; agentic AI merely provides a more powerful medium for exercising it.
Conclusion: Discovery as the Highest‑Value Professional Output
The article closes by reminding readers that “The professionals who thrive will not be the ones who automate the most. They will be the ones who understand that discovery is the highest‑value thing a professional produces, and who have the judgment to direct intelligence toward it deliberately.” By configuring agents with rich context, capable actions, and pointed orientations, and by letting the friction between competing analyses surface hidden insights, professionals can move beyond incremental improvements to genuine strategic breakthroughs. In an era where data is abundant but true insight is scarce, mastering the art of directing intelligence may become the defining competence of effective leadership.
https://sloanreview.mit.edu/article/stop-prompting-ai-start-directing-it/