It began with a single post on LinkedIn. The core observation sounded technical at first glance, but was deeply organizational: Agents versus Skills.
With classic agents, a human decides which specialist to deploy. With skills, dynamic descriptions are ingested, and the model itself chooses what it needs. The power of selection shifts from the employee to the system. Pair that with million-token context windows and autonomous reasoning loops, and you no longer have isolated assistants. You have a single, self-orchestrating swarm.
I read that and thought: This is the future. And immediately after: That is a future that nobody controls.
The Illusion of the Monolith
Every major provider is quietly abandoning standalone models:
-
Anthropic orchestrates dynamic worker swarms that spin up parallel sub-agents.
-
OpenAI routes queries through a multi-model backend.
-
Microsoft Copilot Studio nests orchestrators, child agents, and connected services.
-
Google coordinates multi-agent ecosystems under Agent Mode.
The user sees a simple text response. Under the hood, up to 50 sub-agents may have been launched for a trivial request, burning 15× the token volume, generating non-deterministic handoffs, and creating opaque digital conglomerates. When multiple corporate systems interact, accountability evaporates. Who is liable when Agent A passes a hallucinated directive to Agent B, causing data loss in an ERP?
The industry is locked in a vanity race to build the cleverest black box. Control is not vanishing because people deny responsibility—it vanishes because it gets lost between the systems.
The Answer: The Node and the Accountability Mesh
You cannot police hundreds of ephemeral sub-agents. But you can police the handovers.
In my seventh book, Safe in the Swarm: The Next Level (written in collaboration with DeepSeek), I introduce the operational architecture to bridge this governance vacuum: The Accountability Mesh.
The mesh anchors governance to deterministic Nodes rather than probabilistic models:
-
Deep Context Inspection: Ingress/egress filters enforce strict byte-level schema compliance (e.g., JSON Schema/Pydantic) and block prompt injections before a sub-agent is reached.
-
Action Control, Not Access Control: Sub-agents never hold API keys or credentials. The node evaluates the semantic request, injects credentials server-side from a vault, and strips sensitive metadata.
-
Circuit Breakers: Hard-kill token and cost thresholds terminate runaway loops before financial or infrastructural damage occurs.
-
Federated Human Accountability: You do not need to name every ephemeral sub-agent. You must name the human responsible for the System, the Interface, and the Process.
Code Stops Code. Humans Bear Responsibility.
Governance cannot remain a post-incident checklist. Traditional firewalls watch the perimeter while swarms interact unchecked inside.
Safe in the Swarm: The Next Level is the logical continuation of Human before the Loop (HB4L) and Multivendor Agentic Swarms. It provides CISOs, architects, and business leaders with the concrete blueprints required to build resilient, sovereign, and auditable agent systems.
Complexity is here. The only question left is whether you build the hubs to govern it—or allow your architecture to become an uncontrollable botnet.
