From Connected Agents to Collective Intelligence

This interview evaluation is sponsored by Outshift by Cisco and was written, edited, and revealed in alignment with our Emerj sponsored content guidelines. Learn extra about our thought management and content material creation companies on our Emerj Media Services page.

​

Agentic AI is operating into the identical wall throughout enterprises: brokers drift, impasse, and propagate errors at machine velocity as a result of the foundations for shared that means, shared state, and managed entry aren’t in place.

UC Berkeley researchers analyzed 1,642 actual execution traces throughout seven manufacturing multi-agent frameworks and found failure charges starting from 41% to 86.7% when brokers had to work collectively reasonably than alone. Their taxonomy exhibits the breakdown is structural, not incidental: 41.8% of failures hint to lacking specification and shared governance (the impasse downside), and 36.9% hint to inter-agent misalignment — brokers speaking previous one another on unsuitable assumptions (the semantic drift downside).

The identical physique of analysis demonstrates that when brokers function with out actual coordination, errors are amplified up to 17x versus a single agent working alone; even with centralized checkpoints, amplification nonetheless runs roughly 4.4x.

Governance hasn’t caught up both. The U.S. National Institute of Standards and Technology solely launched its AI Agent Standards Initiative in February 2026, with interoperability steerage not due till This fall 2026 — that means the federal reference framework enterprises govern towards nonetheless doesn’t deal with multi-agent coordination.

Connectivity between brokers exists; collaboration infrastructure doesn’t. Until enterprises construct the shared semantic, reminiscence, and governance layers that this analysis identifies as lacking, multi-agent initiatives will proceed to fail in ways in which are actually well-documented, measurable, and dear.

Emerj’s Yolandi de Weerdt was just lately joined by Guillaume De Saint Marc, VP of Engineering and AI/ML at Outshift by Cisco, to discover how enterprises can transfer from related brokers to coordinated intelligence.

This article examines three insights that make clear why multi‑agent methods stall at scale and what architectural circumstances enable them to function as dependable, collaborative intelligence contained in the enterprise:

• Semantic alignment as the idea of coordinated agent habits: Shared that means and chronic context forestall the drift, impasse, and amnesia that trigger multi‑agent workflows to break the second duties require collaboration.
• Agent‑particular controls as the muse of secure scaling: Upfront safety, observability, and entry governance keep away from the expensive re‑structure that turns into inevitable when brokers fail at machine velocity inside manufacturing workflows.
• Open interoperability as the trail to multi‑agent ecosystem development: Validating one workflow on open foundations creates a scalable anchor that lets future brokers — inner or vendor‑supplied — collaborate with out architectural obstacles.

Listen to the total episode beneath:

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 expertise govt with intensive expertise main engineering, product structure, and rising expertise initiatives. He at the moment serves as Vice President of Engineering at Outshift by Cisco, the place he oversees engineering, software program structure, and platform technique throughout initiatives spanning generative AI, agentic AI, cloud applied sciences, quantum networking, and safety. Prior to his present function, Guillaume held a number of engineering and innovation management positions at Cisco, together with Senior Director of Emerging Technologies & Incubation and Senior Director inside Cisco’s Chief Technology and Architecture Office, main innovation packages, analysis initiatives, and collaborations with startups, universities, and open-source communities. Earlier in his profession, he held govt management roles in R&D, structure, and product administration at NDS and Canal+.

Semantic Alignment because the Basis of Coordinated Agent Behavior

Guillaume De Saint Marc opens the dialog by stating his view of the core failure of multi‑agent methods: brokers don’t fail on the job — they fail on the interpretation. When two brokers learn the identical instruction and derive completely different intent, coordination collapses. And as a result of enterprises don’t have any mechanism to implement shared intent, these divergences compound at machine velocity.

He frames this not as a mannequin weak spot however as a semantic governance hole. Connectivity shouldn’t be coordination; coordination solely emerges when brokers share that means, context, and state. Without a shared semantic layer, each handoff between brokers turns into some extent of divergence — and divergence is what breaks workflows.

To illustrate the operational penalties, Guillaume highlights three failure modes that seem the second duties require collaboration:

• Interpretation drift — brokers progressively diverge of their understanding of the duty.
• Coordination impasse — brokers wait on one another as a result of their inner states not align.
• Context amnesia — brokers lose observe of prior selections, forcing people to intervene.

Where this dialog turns into materially helpful is in how Guillaume defines the semantic layer itself — not as a single artifact, however as a ruled set of shared buildings that each agent should use:

• A shared ontology that defines the objects, actions, and relationships brokers function on.
• A job grammar that standardizes how directions, constraints, and targets are expressed.
• A persistent context retailer that brokers learn from and write to, making certain continuity of state.
• A semantic validator that checks whether or not agent outputs conform to the shared that means mannequin earlier than they propagate.

Guillaume’s viewpoint on why this layer determines whether or not brokers can collaborate:

“If brokers don’t cause from the identical ontology, the identical job grammar, and the identical context, they aren’t collaborating — they’re improvising. And improvisation at machine velocity is chaos. The semantic layer is what forces each agent to function from the identical psychological mannequin, irrespective of who constructed it or the place it runs.”  

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

Taken collectively, Guillaume’s framing exhibits why enterprises evaluating multi‑agent methods should start by understanding how that means, context, and state are shared — or not shared — throughout their workflows.

Agent‑Specific Controls because the Foundation of Safe Scaling

Guillaume shifts from semantics to structure, warning that the stakes are clear: agentic methods fail at machine velocity. Controls aren’t a part‑two concern — they’re the circumstances that make scaling attainable in any respect.

He stresses that brokers should be handled as first‑class actors with identities, privileges, and audit necessities. To make this concrete, Guillaume describes a sample he sees repeatedly: groups deploy brokers with broad entry and minimal observability, and all the pieces works effective in isolation. But as soon as these brokers contact manufacturing methods, a single mis‑permissioned motion forces emergency rollback, guide triage, or a full architectural rebuild.

Guillaume outlines 4 classes of controls that decide whether or not agentic methods scale safely:

• Identity and revocation — brokers will need to have verifiable identities and revocable credentials.
• Semantic observability — leaders want visibility into why an agent acted, not simply what it did.
• Access governance — brokers should function below least‑privilege guidelines enforced constantly.
• Cross‑system interoperability — controls should operate throughout heterogeneous environments, not simply inside a single vendor stack.

Why these controls should be constructed upfront, in accordance to Guillaume:

“The hazard isn’t the error — it’s the velocity of the error. Without id, observability, and entry governance, each error turns into a system‑extensive occasion. Retrofitting controls after that time isn’t a repair. 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 sample Guillaume has seen throughout each enterprise deployment.

His expertise throughout deployments underscores that the reliability of agentic methods relies upon much less on mannequin efficiency than on the power of the controls surrounding them.

Open Interoperability because the Path to Multi‑Agent Ecosystem Growth

The pressure Guillaume surfaces is unavoidable; multi‑agent methods can’t scale inside partitions. When brokers are confined to a single proprietary stack, they inherit its boundaries: information silos, orchestration constraints, and integration bottlenecks. The end result is predictable — brokers that carry out effectively individually however fail to collaborate throughout the enterprise.

Guillaume closes the dialog by addressing this constraint instantly. Vendor lock‑in, legacy orchestration methods, and fragmented information environments make cross‑system coordination unimaginable. His steerage is pragmatic: don’t try a disruptive migration. Instead, validate one actual workflow on open foundations. That workflow turns into the anchor for future brokers — inner or vendor‑supplied — to plug into with out architectural friction.

He additionally outlines the pitfalls he sees repeatedly:

• Closed ecosystems that may’t combine with crucial legacy methods.
• Agents that interpret information otherwise as a result of the underlying semantics are proprietary.
• Pilots that work, however scaling requires re‑architecting each workflow to add a single new agent.

Guillaume’s factors to why open foundations decide lengthy‑time period scalability:

“Closed methods provide you with quick pilots and arduous ceilings. Open foundations provide you with slower pilots and no ceilings in any respect. If you need an ecosystem the place new brokers can be a part of with out breaking what’s already working, openness isn’t a choice — it’s the one viable structure.”  

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

Interoperability isn’t about being open — it’s about avoiding architectural useless ends. Closed methods lure brokers inside the bounds of a single stack; open foundations let enterprises add, change, and scale brokers with out triggering cascading re‑structure.

Guillaume’s perspective highlights that the lengthy‑time period viability of agentic ecosystems is dependent upon whether or not new brokers can be a part of present workflows with out requiring disruptive architectural adjustments.