By 2027, 40% of enterprises will shut down or demote their autonomous AI agents. That is not a prediction from a skeptic. That is Gartner’s May 2026 forecast, published after studying the governance failures enterprises are already discovering in production. The agents are real. The governance layer underneath them is not.
The question is not whether this crisis is coming. It is already here. The question is what the governance layer actually looks like, and why the industry has been naming the wrong problem for the last two years.
The layer underneath them is causal topology: the structural map of how components influence each other under operational load, how that influence propagates, and how the relationship structure shifts over time. Cisco Outshift named this layer publicly when they featured the work DataWell is doing inside the AGNTCY ecosystem.
The Industry Named the Wrong Problem
For two years, the conversation about AI governance has centered on model outputs. Alignment. Guardrails. Safety at the generation layer. The industry assumed that controlling model outputs would control system behavior.
It does not. Agents compose at runtime. Governance happens at the dependency layer, not the generation layer. Forrester said so directly in their 2025 AI CISO report: “No tool today is built to automatically observe an agent’s intent or actions as it spans an organization’s business systems.” The same report drew the distinction clearly: the problem is not “classic IT observability into whether something happened” but rather “contextual visibility into why something happened.”
Logs tell you what happened. Dashboards tell you what each component reported. Alerts tell you something crossed a threshold. None of them map why the system behaved the way it did, which components drove which outcomes, or where pressure is building before it becomes an incident.
The result: according to Splunk and Oxford Economics’ 2026 research, unplanned downtime now costs the Global 2000 $600 billion annually, up 50% in just two years. For the average organization, that is $15,000 lost every single minute a critical system is down. Only 38% of technology executives say they always find root cause, and 47% admit their customers detect the outage before their own monitoring tools do. The agents are compounding a problem that monitoring was already failing to solve.
What Causal Topology Actually Is
Causal topology is the map underneath your infrastructure. Not a map of what each component reported. A map of how components actually influence each other under operational load, how that influence propagates across multi-hop dependency chains, and how the relationship structure shifts over time.
The difference between causal topology and conventional observability is the difference between knowing a patient’s vitals and knowing how their organs function together. Vitals tell you something is wrong. Causal topology tells you which system drove the failure and what it would take to prevent the next one.
This is what a causal topology engine produces. In a real infrastructure analysis, the engine identified 211 directed relationships across a single 8-hour operational window. It mapped how system drivers propagated influence across multi-hop dependency chains, tracked how the relationship topology shifted under different operational loads, and surfaced high-confidence influence pathways while the conditions were still visible in the data, before they became incidents.
In one analysis, the engine identified a 70.5% improvement in disk I/O wait time available through a single targeted change. Not because an alert fired. Because the causal map showed where the pressure was concentrated before anyone felt it.
The difference between looking at metrics and understanding system behavior. Not more dashboards. Not louder alerts. A structural map of the dependency relationships underneath the metrics, produced from operational telemetry, readable by engineering teams.
This is Decision Trust: the infrastructure principle that you cannot govern what you cannot causally understand. Logging captures what happened. Causal topology maps why it happened.
Why Cisco Named It
When AGNTCY, the Linux Foundation project backed by Cisco, Google, Red Hat, Dell, Oracle, and more than 80 partner organizations, built the connective infrastructure for the Internet of Agents, they solved the right first problem: identity, directory, secure messaging, and interoperability. Without those rails, agent cooperation across organizational boundaries cannot happen.
But once the rails are in place and agents are cooperating, a new question begins. Are those agents still behaving within bounds? Are the dependency relationships between components shifting in ways nobody has flagged? Is pressure building somewhere in the system that nobody has named yet?
Those are not questions about identity. They are questions about causal topology. This is the layer Cisco Outshift identified when they featured the causal topology work happening inside AGNTCY. As Noelle Flint wrote: “This is not metrics. It’s system intelligence.”
AGNTCY answers who an agent is and how it connects. Causal topology maps what happens after the connection opens. The behavior. The dependencies. The pressure building somewhere in the chain before anyone names it.
Agent Zero: The On-Ramp
To bring causal topology into AGNTCY-connected environments, a lightweight intelligence agent enters the ecosystem credentialed through AGNTCY Identity, communicates over SLIM, and registers in the agent directory. Its role: invite other agents to consent to telemetry sharing, then produce a Behavioral Confidence Profile for the system they form together: where the dependency topology is stable, where behavior is shifting, where pressure is building, and what the confidence level is on each finding.
That profile is readable by an engineer who wants the technical detail and by an executive who needs the summary in under thirty seconds. The governance layer is not a separate audit process. It is a live structural map, produced continuously from the system’s own telemetry.
You can read the full technical documentation on how DataWell integrates with the AGNTCY ecosystem at getdatawell.com/blog/agntcy-datawell-internet-of-agents.
The Window Is Closing
The EU AI Act’s transparency and high-risk obligations take enforcement effect in August 2026. Non-compliance penalties reach up to 15 million euros or 3% of global turnover for systems that cannot demonstrate traceability. The agents already in production across most enterprises cannot demonstrate it today.
Gartner’s 40% decommission projection is not a warning about AI adoption. It is a warning about what happens when enterprises discover, after a production incident, that they deployed agents into systems they could not actually see. The governance gap is not a governance problem. It is a topology problem.
What has been missing is not more monitoring. Not louder alerts. Not guardrails on systems nobody has fully traced. A causal topology engine that produces a structural readout of how systems actually behave, built from operational telemetry, available before the incident.
Early Access Is Open
Causal topology infrastructure is live and in production across fintech, cybersecurity, predictive maintenance, and enterprise AI environments where failure has real consequences. Organizations running autonomous agents without the ability to trace causal dependency structures underneath them are operating with a governance gap that no existing observability tool was built to close. The infrastructure to close it is available at getdatawell.com.
If your organization is running autonomous agents in production and cannot trace the causal structure underneath them, you are operating blind at scale. Early access to the causal topology engine is available through the site for organizations where that is not acceptable.
The governance layer Gartner predicted is not coming. It is here. And it is built on causal topology.
Benjamin Torres is Founder and CEO of Versai Labs and creator of DataWell. Versai Labs architects system intelligence for organizations where failure has real consequences. Learn more at versailabs.com.