In our first article, we discussed the arrival of the agentic ecosystem and the need for design due to its impact on how decisions are made, actions are taken and accountability is distributed.
The next challenge is less visible, but equally impactful — and more consequential.
Agentic systems do not behave like traditional software, nor do they operate according to defined project plans. They move through organizations as currents rather than commands, invoking tools, negotiating with other agents, and rewriting instructions in real time. Much of this activity occurs below the surfaces designed for human awareness. Dashboards show outcomes. They do not reveal the pathways taken to reach the outcomes and the multitude of decisions and information processed. In the past, humans controlled both the pathways taken and to a large degree the information processed. That has all changed.
Individuals and teams are beginning to sense that something fundamental has changed. Decisions appear to materialize rather than originate. Work advances without clear authorship. Optimization unfolds continuously, often in directions no one explicitly chose. Organizations are not built for these types of results given that most organizational structures are a version of the industrial revolution of the late 1900s.
Recent events have some individuals aware of this challenge. Enterprises experimenting with autonomous coding environments and operational agents are discovering that pilots designed to be self-contained are quietly interacting with live infrastructure. Optimization routines pursuing efficiency targets are conflicting with security constraints. Agents tasked with solving local problems are improvising global solutions.
While specificity was designed, multiplicity and interconnection emerged.
This is the visibility gap.
Recent “sandboxed” deployments have shown how porous and challenging containment can be. Systems running on separate machines are still able to reach sensitive data through shared credentials, identity linkages, or connected services. It is not that agents are actively trying to bypass safeguards, they are just doing what they are designed to do – fulfill their objectives – and in doing so inadvertently jump their guardrails. These systems are not reckless. They are literal. They act on the world as it is, not as designers assume it to be or mean for the agents to behave.
In several recent enterprise pilots, agents designed to synthesize information across systems surfaced restricted data because they treated accessibility as permission. The boundary between “can retrieve” and “should retrieve” proved thinner than anticipated.
At the same time, a growing number of senior researchers have stepped away from leading labs, citing concerns that deployment pace is outstripping safety readiness. Their departures reflect a broader tension between innovation pressure, institutional preparedness and the broader systemic impacts.
As we mentioned in our previous article, small shifts accumulate. Context erodes as agents interact with each other, and optimization routines interact. Authority diffuses across chains of delegation and what hat began as assistance can evolve into autonomous decision-making without a clear moment of transition.
Nothing breaks, no warning is given, the system simply becomes something else.
Every organization is feeling the pressure to innovate, and as a result timelines are reduced. The pace of deployment intensifies the challenge. Competitive pressure compresses the cycle from experimentation to production. Capabilities once evaluated over years now enter operational environments in months or weeks. Governance structures designed to take in multiple disciplines, cannot keep up with the continuous change and cannot determine the needed guardrails fast enough. We are moving from software that executes instructions to ecosystems that pursue objectives.
We laid out the case in our previous article for supervision as part of the design process. We believe that it almost must be I embedded into the architecture itself.
Governance cannot be just a system of policies, controls, training and overshight, it must include architectural requirements as well as deployment rules. Some of the most essential questions are architectural: Where must human judgment remain decisive? How should authority degrade across chains of agents? Which decisions require explicit confirmation before execution? What signals indicate that the system is drifting from its intended purpose? These questions determine whether an ecosystem remains legible to those responsible for it.
The discovery phase requires more than documentation, it requires situational awareness. And that situational awareness is multi-disciplinary. And that awareness needs to be shared across teams in order to deploy agents well.
Mapping an agentic ecosystem often reveals a landscape more complex than anticipated. Agents accumulate capabilities through integrations. External services become embedded dependencies. Data flows extend beyond original boundaries. Organizational ownership becomes diffuse as responsibility spans technical, operational, and business domains.
No single team holds the full picture.
This fragmentation introduces a new kind of risk, not technological, but institutional. When systems cross organizational boundaries faster than governance can adapt, accountability becomes performative rather than operational. Everyone is responsible in theory. No one is responsible in practice and yet everyone recognizes the importance of owning the outcomes.
At the same time, the language available to leaders has not caught up to the experience. Terms like governance, enablement, or upskilling imply incremental adjustment. What leaders describe instead is velocity and compression, the sense of managing forces that accelerate as they are applied.
This is not a failure of preparation. It is a failure of frameworks and a reflection of the enormity of the change and of the resulting challenge.
Discovery becomes the first act of responsible leadership in the agentic era. Before organizations can design supervision, they must understand the ecosystem they have already set in motion, intentionally or otherwise. This requires mapping agents, interactions, authority structures, data pathways, and optimization loops as they actually function. (We will address this leadership challenge and the resulting need for more training and expertise in a future article.)
Visibility precedes control.
Without it, safety is shaky and fragmented. Accountability is ambiguous. Risk is ambient, present everywhere and owned nowhere.
The organizations that navigate this transition successfully will not be those that deploy agents fastest. They will be those that maintain the ability to understand and guide the systems they create.
In our next article, we will move from discovery to design, exploring how ownership and supervision can be architected into the process so that safety, accountability and performance scale together rather than compete. A difficult but necessary step-change.
The defining capability of the next decade will not be autonomy. It will be the ability to remain oriented within ever-increasing autonomous systems.