Published on 00/00/0000
Last updated on 00/00/0000
Published on 00/00/0000
Last updated on 00/00/0000
Share
Share
INSIDE OUTSHIFT
9 min read
Share
For decades, enterprise production environments have been built on deterministic principles—predictable inputs produce predictable outputs. This foundational certainty has underpinned everything from basic calculators to modern data centers.
But AI agents are different beasts entirely. They're probabilistic, autonomous, and increasingly trusted with sensitive operations across organizational boundaries. Think of it like the difference between a programmed assembly line robot and a human consultant who shows up, assesses the situation, and makes judgment calls.
In OPAQUE’s AI Confidential podcast, Vijoy Pandey, Senior Vice President and General Manager at Outshift by Cisco, tackled the messy realities of securing these autonomous agents.
The conversation with hosts Aaron Fulkerson and Matt Hinkle cuts through the hype to address practical questions: How do you verify an agent's identity when it's operating at machine speed? What happens when agents from different organizations need to collaborate? And why do we need an entirely new internet just for AI?
Here are some key insights from their discussion:
Probabilistic computing changes everything. Outcomes are no longer guaranteed; advanced AI systems operate at machine speed and scale with the unpredictability of human-like behavior. Simply put, an agent, given the same input, might produce different, yet equally valid outcomes.
“Humans are probabilistic in nature, but they’re [AI agents] operating at machine speed and scale, and we have not seen that happen before,” Pandey says.
Because of this uncertainty, enterprises need a new approach. They will need to rethink how they design, secure, and interact with systems. Agents may come from diverse origins, built by different vendors, and using various frameworks. They might also operate across multiple clouds. This heterogeneity adds complexity to enterprise systems.
It's a shift that challenges the very intuition about digital predictability. It forces an embrace of a world where certainty is replaced by probabilities, and control must adapt to autonomy.
With agents operating probabilistically and at machine speed, identity and trust issues come into play. Traditional security frameworks, designed for either human users or static machine services, no longer suffice for these hybrid entities. A new approach for authentication and authorization is required.
Task-based identities offer a solution. They grant agents authorization to perform specific, time-bound tasks with strictly controlled access. Once the task is complete, the agent reverts to default state, minimizing risk. This granular approach ensures an agent’s access and capabilities are precisely aligned with its current objective.
To implement this, Outshift has launched AI Agent Identity through the AGNTCY. The framework assigns, verifies, and manages credentials for AI agents.
The W3C’s decentralized identity platform, complaint with decentralized identifiers (DID) standards, aims to prevent any single entity from controlling agent identity. This approach fosters a truly open and trustworthy Internet of Agents.
“If we are true to our word, the Internet of Agents should be open and interoperable,” Pandey says.
Just as CP, UDP, QUIC over IP, plus HTTP, DNS, and other essential protocols, laid the groundwork for the original internet and HTTPS became the transport layer for the cloud era, the Internet of Agents requires a new communication paradigm. Agents don't just exchange data, they need to understand context, intent, and capabilities in a probabilistic world.
Pandey believes that there will be two additional layers above HTTPs: semantic and syntactic.
The semantic layer would address the meaning of agent communication—interpreting what an agent truly intends. It will consider context and confidence levels. On the other hand, the syntactic layer would focus on the underlying frameworks and models used by agents, ensuring compatibility and interoperability among agents.
“Given Cisco's deep heritage in networking and defining communication protocols, its leadership in developing these new layers for agent-to-agent interaction is a natural and necessary evolution. These layers will enable true collaboration and build trust in a world where machines are increasingly conversing with each other,” Pandey says.
One of the use cases Outshift is developing is a site reliability engineering (SRE) agent named SRE Jarvis. SRE Jarvis is a multi-agent system, composed of over 20 agents and 50 tools and data backends designed to automate complex workflows in the SRE domain.
Using open source components, SRE Javis integrates with IT infrastructure like Kubernetes, PagerDuty, and AWS to help engineers automate self-serve tasks, such as bootstrapping CI/CD pipelines, granting access to approved large language models (LLMs), and setting up EC2 developer instances.
SRE Jarvis also incorporates advanced capabilities like observability across systems, identity management, and network deployment debugging. Antically routed workflows allow human oversight in complex or non-deterministic scenarios, bridging the gap between deterministic and probabilistic systems.
While the ultimate vision of fully self-assembling, self-forming agents is still years away, initiatives like SRE Jarvis prove that agentic AI is already delivering value by tackling blockers and boosting productivity in IT operations.
While agentic AI is all the buzz, most enterprises are still in search of a scalable implementation of their products and workflow.
For a deeper understanding of this emerging landscape, listen to the complete AI Confidential podcast with Pandey, available on Spotify, Apple Podcasts, and YouTube.
For decades, enterprise production environments have been built on deterministic principles—predictable inputs produce predictable outputs. This foundational certainty has underpinned everything from basic calculators to modern data centers.
But AI agents are different beasts entirely. They're probabilistic, autonomous, and increasingly trusted with sensitive operations across organizational boundaries. Think of it like the difference between a programmed assembly line robot and a human consultant who shows up, assesses the situation, and makes judgment calls.
In OPAQUE’s AI Confidential podcast, Vijoy Pandey, Senior Vice President and General Manager at Outshift by Cisco, tackled the messy realities of securing these autonomous agents.
The conversation with hosts Aaron Fulkerson and Matt Hinkle cuts through the hype to address practical questions: How do you verify an agent's identity when it's operating at machine speed? What happens when agents from different organizations need to collaborate? And why do we need an entirely new internet just for AI?
Here are some key insights from their discussion:
Probabilistic computing changes everything. Outcomes are no longer guaranteed; advanced AI systems operate at machine speed and scale with the unpredictability of human-like behavior. Simply put, an agent, given the same input, might produce different, yet equally valid outcomes.
“Humans are probabilistic in nature, but they’re [AI agents] operating at machine speed and scale, and we have not seen that happen before,” Pandey says.
Because of this uncertainty, enterprises need a new approach. They will need to rethink how they design, secure, and interact with systems. Agents may come from diverse origins, built by different vendors, and using various frameworks. They might also operate across multiple clouds. This heterogeneity adds complexity to enterprise systems.
It's a shift that challenges the very intuition about digital predictability. It forces an embrace of a world where certainty is replaced by probabilities, and control must adapt to autonomy.
With agents operating probabilistically and at machine speed, identity and trust issues come into play. Traditional security frameworks, designed for either human users or static machine services, no longer suffice for these hybrid entities. A new approach for authentication and authorization is required.
Task-based identities offer a solution. They grant agents authorization to perform specific, time-bound tasks with strictly controlled access. Once the task is complete, the agent reverts to default state, minimizing risk. This granular approach ensures an agent’s access and capabilities are precisely aligned with its current objective.
To implement this, Outshift has launched AI Agent Identity through the AGNTCY. The framework assigns, verifies, and manages credentials for AI agents.
The W3C’s decentralized identity platform, complaint with decentralized identifiers (DID) standards, aims to prevent any single entity from controlling agent identity. This approach fosters a truly open and trustworthy Internet of Agents.
“If we are true to our word, the Internet of Agents should be open and interoperable,” Pandey says.
Just as CP, UDP, QUIC over IP, plus HTTP, DNS, and other essential protocols, laid the groundwork for the original internet and HTTPS became the transport layer for the cloud era, the Internet of Agents requires a new communication paradigm. Agents don't just exchange data, they need to understand context, intent, and capabilities in a probabilistic world.
Pandey believes that there will be two additional layers above HTTPs: semantic and syntactic.
The semantic layer would address the meaning of agent communication—interpreting what an agent truly intends. It will consider context and confidence levels. On the other hand, the syntactic layer would focus on the underlying frameworks and models used by agents, ensuring compatibility and interoperability among agents.
“Given Cisco's deep heritage in networking and defining communication protocols, its leadership in developing these new layers for agent-to-agent interaction is a natural and necessary evolution. These layers will enable true collaboration and build trust in a world where machines are increasingly conversing with each other,” Pandey says.
One of the use cases Outshift is developing is a site reliability engineering (SRE) agent named SRE Jarvis. SRE Jarvis is a multi-agent system, composed of over 20 agents and 50 tools and data backends designed to automate complex workflows in the SRE domain.
Using open source components, SRE Javis integrates with IT infrastructure like Kubernetes, PagerDuty, and AWS to help engineers automate self-serve tasks, such as bootstrapping CI/CD pipelines, granting access to approved large language models (LLMs), and setting up EC2 developer instances.
SRE Jarvis also incorporates advanced capabilities like observability across systems, identity management, and network deployment debugging. Antically routed workflows allow human oversight in complex or non-deterministic scenarios, bridging the gap between deterministic and probabilistic systems.
While the ultimate vision of fully self-assembling, self-forming agents is still years away, initiatives like SRE Jarvis prove that agentic AI is already delivering value by tackling blockers and boosting productivity in IT operations.
While agentic AI is all the buzz, most enterprises are still in search of a scalable implementation of their products and workflow.
For a deeper understanding of this emerging landscape, listen to the complete AI Confidential podcast with Pandey, available on Spotify, Apple Podcasts, and YouTube.
Get emerging insights on innovative technology straight to your inbox.
Outshift is leading the way in building an open, interoperable, agent-first, quantum-safe infrastructure for the future of artificial intelligence.
* No email required
The Shift is Outshift’s exclusive newsletter.
Get the latest news and updates on agentic AI, quantum, next-gen infra, and other groundbreaking innovations shaping the future of technology straight to your inbox.
