How AI Agent Swarms Might Be AI's Next Leap — With Guillaume de Saint-Marc
TL;DR
Maltbook showed both the promise and the emptiness of current agent swarms — Guillaume de Saint-Marc said the viral network of 1M+ agents, 14M comments, and 2.3M posts proved agents can find and message each other at scale, but it was mostly a “theater of collaboration” with no shared state, governance, or real coordination.
Cisco’s thesis is that AI advances through horizontal scaling, not just smarter single agents — Outshift has spent more than two years building for many agents working together like distributed systems, arguing that real capability comes from swarms that self-form around a problem rather than fixed step-by-step workflows.
The hard part is not connectivity anymore — it’s cognition — After launching its open-source “internet of agents” stack via AGNTCY with Google, Oracle, Red Hat, and Dell, the team found that connecting, identifying, discovering, and observing agents is necessary but insufficient for complex missions.
Guillaume proposes a new ‘layer 9’ semantic layer for agent coordination — Beyond cloud-native infrastructure and protocol-level agent communication like A2A and MCP, he says swarms need message-level understanding of intent, delegation, and knowledge-sharing so agents can actually sync and reason together.
Enterprise-safe swarms depend on guardrails like rooms, revocation, and semantic tool checks — AGNTCY uses Slack-like agent rooms, MLS-encrypted communication, cryptographically signed agent cards, and ‘TBAC’ so a micro-agent can reject nonsense behavior like an agent trying to make a transaction when it was only asked to check FX rates.
The concrete use case is IT crisis response collapsing days of work into minutes — In a severe outage, Guillaume imagines SRE, security, observability, root-cause, and crisis-communications agents from different vendors coordinating in real time, with early experiments cutting root-cause analysis from three days of experts in a room to a few minutes.
The Breakdown
Maltbook Was Wild — and Also Kind of Fake
Alex opens with the perfect hook: Maltbook, the agent social network that looked like agents had built their own online society, maybe even a religion. Guillaume says everyone was “eating popcorn” watching it go viral, but his engineering takeaway was sharper: the spectacle was real, the collaboration wasn’t. Agents could connect at scale, but without shared state, governance, or meaningful coordination, it was mostly pattern-matching social behavior — plus a mess of security holes and stolen credentials.
Cisco’s Bet: Intelligence Comes From Swarms, Not Solo Agents
Guillaume says Outshift has been focused on this for more than two years with a simple thesis: the next leap comes from “horizontal scaling,” meaning lots of agents working together at machine speed. That led to AGNTCY, Cisco’s open-source “internet of agents” effort, built around four basics: connect agents, give them identity, help them discover each other, and observe what they do. It’s a very Cisco framing — connect, secure, observe — but applied to agents as a new kind of entity that behaves like both software workload and user.
Why Workflows Aren’t Enough Anymore
He draws a line between replacing steps in a predefined workflow with agents — useful, but limited — and something much more interesting: self-forming collaboration. In his example, reasoning itself can spawn sub-agents as needed, similar to what people see in systems like OpenAI’s operator-style loops. The big unlock is that agents can discover other agents with the right skills in real time and pull them into a mission, instead of following a prewritten script.
The Safety Problem: How Do You Stop a Runaway Swarm?
Alex pushes on the obvious fear: if swarms self-form, how do you keep them from going off the rails? Guillaume’s answer is a stack of controls: strict connectivity rules, Slack/Webex-style “rooms” for scoped collaboration, MLS-encrypted transport, and the ability to revoke one rogue agent without breaking everyone else. On top of that is identity via signed “agent cards” and TBAC, which limits what tools an agent can use — including semantic checks so an agent asked to look up a currency rate can’t suddenly justify making a transaction.
From Layer 7 to ‘Layer 9’ — The Internet Needs New Agent Plumbing
This is the most technical turn in the conversation. Guillaume says the classic stack stops too low: today’s internet and cloud-native systems get you to application-level communication, but agents need two more layers. He calls one the syntactic layer, where protocols like A2A and MCP live, and the next the semantic layer — “layer 9” — where the system understands whether a message is sharing intent, knowledge, or a delegated task.
The Real Prize: Multi-Agent Incident Response
When Alex asks what swarms can do that one agent can’t, Guillaume gets concrete fast: resolving a serious IT outage. That problem needs SRE agents, security agents, observability agents, root-cause analysis agents, and even crisis-communications agents, often from different vendors like Splunk or Cisco. He says early experiments have already shown root-cause work dropping from three days of human experts in a room to just minutes, though he’s careful to say they’re “not fully there yet.”
Why Open Source So Much of This?
Alex asks the obvious business question: why would Cisco give away infrastructure that could be a competitive edge? Guillaume’s answer is that agent interoperability is too foundational to live inside walled gardens, just like the internet itself couldn’t have worked that way. He says more than 80% of what they build should be open because no single company can create the “internet of cognition” alone — and Cisco can still differentiate on product velocity.
What’s Available Now vs. What’s Coming Next
Guillaume closes by splitting the roadmap in two. The internet-of-agents layer — identity, connectivity, discovery, observability — is already out, in production paths, and tied closely to A2A, with open-source code and examples like “coffee AGNTCY.” The newer “internet of cognition” work is still early, but white papers are out, code was slated to begin landing in April, and the team is systematically testing where swarms derail, mapping seven or eight recurring cognition failures they want the architecture to solve.