From Connected Agents to Collective Intelligence

This interview analysis is sponsored by Outshift by Cisco and was written, edited, and published in alignment with our Emerj sponsored content guidelines. Learn more about our thought leadership and content creation services on our Emerj Media Services page.

​

Agentic AI is running into the same wall across enterprises: agents drift, deadlock, and propagate errors at machine speed because the foundations for shared meaning, shared state, and controlled access aren’t in place.

UC Berkeley researchers analyzed 1,642 real execution traces across seven production multi-agent frameworks and found failure rates ranging from 41% to 86.7% when agents had to work together rather than alone. Their taxonomy shows the breakdown is structural, not incidental: 41.8% of failures trace to missing specification and shared governance (the deadlock problem), and 36.9% trace to inter-agent misalignment — agents talking past each other on wrong assumptions (the semantic drift problem).

The same body of research demonstrates that when agents operate without real coordination, errors are amplified up to 17x versus a single agent working alone; even with centralized checkpoints, amplification still runs roughly 4.4x.

Governance hasn’t caught up either. The U.S. National Institute of Standards and Technology only launched its AI Agent Standards Initiative in February 2026, with interoperability guidance not due until Q4 2026 — meaning the federal reference framework enterprises govern against still doesn’t address multi-agent coordination.

Connectivity between agents exists; collaboration infrastructure does not. Until enterprises build the shared semantic, memory, and governance layers that this research identifies as missing, multi-agent initiatives will continue to fail in ways that are now well-documented, measurable, and costly.

Emerj’s Yolandi de Weerdt was recently joined by Guillaume De Saint Marc, VP of Engineering and AI/ML at Outshift by Cisco, to explore how enterprises can move from connected agents to coordinated intelligence.

This article examines three insights that clarify why multi‑agent systems stall at scale and what architectural conditions allow them to operate as reliable, collaborative intelligence inside the enterprise:

• Semantic alignment as the basis of coordinated agent behavior: Shared meaning and persistent context prevent the drift, deadlock, and amnesia that cause multi‑agent workflows to break the moment tasks require collaboration.
• Agent‑specific controls as the foundation of safe scaling: Upfront security, observability, and access governance avoid the costly re‑architecture that becomes inevitable when agents fail at machine speed inside production workflows.
• Open interoperability as the path to multi‑agent ecosystem growth: Validating one workflow on open foundations creates a scalable anchor that lets future agents — internal or vendor‑provided — collaborate without architectural barriers.

Listen to the full episode below:

Episode:  From Connected Agents to Collective Intelligence with Guillaume De Saint Marc of Outshift by Cisco

​Guest: Guillaume De Saint Marc, VP Engineering and AI/ML, Outshift by Cisco

Expertise: Generative AI, Agentic AI, Cloud-Native Architecture, Emerging Technologies

Brief Recognition: Guillaume De Saint Marc is a technology executive with extensive experience leading engineering, product architecture, and emerging technology initiatives. He currently serves as Vice President of Engineering at Outshift by Cisco, where he oversees engineering, software architecture, and platform strategy across projects spanning generative AI, agentic AI, cloud technologies, quantum networking, and security. Prior to his current role, Guillaume held several engineering and innovation leadership positions at Cisco, including Senior Director of Emerging Technologies & Incubation and Senior Director within Cisco’s Chief Technology and Architecture Office, leading innovation programs, research initiatives, and collaborations with startups, universities, and open-source communities. Earlier in his career, he held executive leadership roles in R&D, architecture, and product management at NDS and Canal+.

Semantic Alignment as the Basis of Coordinated Agent Behavior

Guillaume De Saint Marc opens the conversation by stating his view of the core failure of multi‑agent systems: agents don’t fail at the task — they fail at the interpretation. When two agents read the same instruction and derive different intent, coordination collapses. And because enterprises have no mechanism to enforce shared intent, these divergences compound at machine speed.

He frames this not as a model weakness but as a semantic governance gap. Connectivity is not coordination; coordination only emerges when agents share meaning, context, and state. Without a shared semantic layer, every handoff between agents becomes a point of divergence — and divergence is what breaks workflows.

To illustrate the operational consequences, Guillaume highlights three failure modes that appear the moment tasks require collaboration:

• Interpretation drift — agents gradually diverge in their understanding of the task.
• Coordination deadlock — agents wait on each other because their internal states no longer align.
• Context amnesia — agents lose track of prior decisions, forcing humans to intervene.

Where this conversation becomes materially useful is in how Guillaume defines the semantic layer itself — not as a single artifact, but as a governed set of shared structures that every agent must use:

• A shared ontology that defines the objects, actions, and relationships agents operate on.
• A task grammar that standardizes how instructions, constraints, and goals are expressed.
• A persistent context store that agents read from and write to, ensuring continuity of state.
• A semantic validator that checks whether agent outputs conform to the shared meaning model before they propagate.

Guillaume’s point of view on why this layer determines whether agents can collaborate:

“If agents don’t reason from the same ontology, the same task grammar, and the same context, they aren’t collaborating — they’re improvising. And improvisation at machine speed is chaos. The semantic layer is what forces every agent to operate from the same mental model, no matter who built it or where it runs.”  

— Guillaume De Saint Marc, VP Engineering and AI/ML, Outshift by Cisco

Taken together, Guillaume’s framing shows why enterprises evaluating multi‑agent systems must begin by understanding how meaning, context, and state are shared — or not shared — across their workflows.

Agent‑Specific Controls as the Foundation of Safe Scaling

Guillaume shifts from semantics to architecture, warning that the stakes are clear: agentic systems fail at machine speed. Controls aren’t a phase‑two concern — they’re the conditions that make scaling possible at all.

He stresses that agents must be treated as first‑class actors with identities, privileges, and audit requirements. To make this concrete, Guillaume describes a pattern he sees repeatedly: teams deploy agents with broad access and minimal observability, and everything works fine in isolation. But once those agents touch production systems, a single mis‑permissioned action forces emergency rollback, manual triage, or a full architectural rebuild.

Guillaume outlines four categories of controls that determine whether agentic systems scale safely:

• Identity and revocation — agents must have verifiable identities and revocable credentials.
• Semantic observability — leaders need visibility into why an agent acted, not just what it did.
• Access governance — agents must operate under least‑privilege rules enforced continuously.
• Cross‑system interoperability — controls must function across heterogeneous environments, not just within a single vendor stack.

Why these controls must be built upfront, according to Guillaume:

“The danger isn’t the mistake — it’s the speed of the mistake. Without identity, observability, and access governance, every error becomes a system‑wide event. Retrofitting controls after that point isn’t a fix. It’s a rebuild.”  

— Guillaume De Saint Marc, VP Engineering and AI/ML, Outshift by Cisco

Agentic AI scales safely when controls come first, not as an afterthought — a pattern Guillaume has seen across every enterprise deployment.

His experience across deployments underscores that the reliability of agentic systems depends less on model performance than on the strength of the controls surrounding them.

Open Interoperability as the Path to Multi‑Agent Ecosystem Growth

The tension Guillaume surfaces is unavoidable; multi‑agent systems can’t scale inside walls. When agents are confined to a single proprietary stack, they inherit its boundaries: data silos, orchestration constraints, and integration bottlenecks. The outcome is predictable — agents that perform well individually but fail to collaborate across the enterprise.

Guillaume closes the conversation by addressing this constraint directly. Vendor lock‑in, legacy orchestration systems, and fragmented data environments make cross‑system coordination impossible. His guidance is pragmatic: don’t attempt a disruptive migration. Instead, validate one real workflow on open foundations. That workflow becomes the anchor for future agents — internal or vendor‑provided — to plug into without architectural friction.

He also outlines the pitfalls he sees repeatedly:

• Closed ecosystems that can’t integrate with critical legacy systems.
• Agents that interpret data differently because the underlying semantics are proprietary.
• Pilots that work, but scaling requires re‑architecting every workflow to add a single new agent.

Guillaume’s points to why open foundations determine long‑term scalability:

“Closed systems give you fast pilots and hard ceilings. Open foundations give you slower pilots and no ceilings at all. If you want an ecosystem where new agents can join without breaking what’s already working, openness isn’t a preference — it’s the only viable architecture.”  

— Guillaume De Saint Marc, VP Engineering and AI/ML, Outshift by Cisco

Interoperability isn’t about being open — it’s about avoiding architectural dead ends. Closed systems trap agents inside the limits of a single stack; open foundations let enterprises add, replace, and scale agents without triggering cascading re‑architecture.

Guillaume’s perspective highlights that the long‑term viability of agentic ecosystems depends on whether new agents can join existing workflows without requiring disruptive architectural changes.