Original source: Outshift by Cisco
This video from Outshift by Cisco covered a lot of ground. 11 segments stood out as worth your time. Everything below links directly to the timestamp in the original video.
Imagine a world where critical IT issues are automatically triaged and reported within seconds after a system outage, freeing up human engineers for more strategic tasks. This technology aims to make that a reality, transforming how organizations manage their incident response.
AI Agent Automates Post-Outage Ticket Management
A sophisticated AI agent supervisor can automatically generate a to-do list and execute parallel tasks to manage post-outage service request (SR) tickets. This deep agent system queries platforms like PagerDuty to identify on-call personnel and Jira for assigned issues, then compiles a comprehensive report in approximately 40 seconds. This capability significantly reduces the manual effort required after system incidents, allowing human on-call staff more rest while ensuring critical issues are tracked.
The system's ability to plan and execute tasks with multiple large language model (LLM) calls demonstrates a notable advancement in autonomous AI, moving beyond simple queries to orchestrate complex operational workflows. By quickly gathering and presenting information on assigned issues, the agent provides a rapid overview that facilitates follow-up actions or reassignment, thereby improving incident response efficiency and potentially reducing mean time to recovery.
"The deep agent supervisor does is, 'how do I go about and plan this?' So it makes a to-do list saying, 'hey I need to go and query PagerDuty in order to find out who was on call, then I need to go and check the Jira, and then I need to put it in a useful format.'"
Outshift Achieves Significant Efficiency Gains with AI Agent System
Outshift has dramatically reduced its dedicated service request personnel from four engineers to less than one through the implementation of an AI agentic system. This system provides 24/7 availability for answering technical questions, automates 30% of service requests end-to-end, and cuts resolution times from half a day to mere minutes. Engineers can now dedicate more time to creative work and automation improvements, enhancing overall productivity.
Beyond routine request automation, the agentic system excels at troubleshooting by rapidly sifting through vast amounts of data from disparate sources at machine speed. This capability significantly reduces the mean time to recover from issues by quickly identifying potential problem areas, a critical advantage in complex cloud-native environments. The success highlights the transformative potential of AI agents in streamlining IT operations and improving service delivery.
"We've been able to reduce that to less than one engineer, and what we have done is like use that time in order to invest on more creative work, improving the automation, improving the tooling."
Future Workforce to Blend Human Oversight with Autonomous AI Agents
The future workforce is envisioned as a hybrid collaboration between humans and autonomous AI agents, with humans providing the strategic direction and agents executing the tactical 'how.' In this model, human roles will focus on product vision, prioritization, ethical considerations, policy setting, and final decision-making. AI agents, on the other hand, will handle execution with high quality and speed, enabling faster product development cycles and quicker market capture.
Successful adoption of this hybrid workforce necessitates robust platform engineering and integrated security measures from the outset. Ensuring repeatability, consistency, and addressing new attack vectors introduced by autonomous systems are paramount. This approach promises enhanced product outcomes, reduced costs, and accelerated time-to-market, fundamentally reshaping how organizations operate and innovate.
"The future workforce is going to be hybrid in terms of humans and these autonomous agents. You're going to have humans providing product vision, prioritization, ethics, policy, and final decisions, and then agents will mostly execute the 'how.'"
AI Agent Streamlines Development Machine Provisioning via Jira
An AI agentic system can fully automate the process of provisioning development machines requested through Jira tickets, significantly reducing manual effort and wait times. When a user submits a request for a development machine, the agent clarifies their intent—for example, offering choices between an EC2 instance or a Kubernetes cluster—before proceeding. The system then automates the VM creation, including raising a pull request for human approval, and finally provides the user with access to the machine and its private key.
This end-to-end automation, which allows for back-and-forth interaction within Jira until the user is satisfied, represents a substantial improvement over traditional, time-consuming manual processes. By understanding natural language requests, automating infrastructure provisioning, and integrating a human approval step, the agentic system enhances developer experience and operational efficiency, transforming routine IT support tasks into a seamless, automated workflow.
"In your GitHubs workflow, it's saying, 'hey I need to bring a human in for approval and I've now raised a PR,' and that's been sent to a human so they could just approve the review and approve the request."
AI Multi-Agent System Boosts Platform Engineering Efficiency at Outshift
Outshift has successfully implemented a multi-agentic AI system to manage the growing complexity of its cloud-native platform engineering, addressing challenges faced by small teams. This AI-powered solution enhances sustainability, reduces repetitive tasks, and improves development velocity. Recognizing its internal success, Outshift open-sourced the system under the name Canoe, contributing to the broader Agency Internet of Agents stack.
The adoption of AI agents has been crucial for Outshift's incubation unit, which struggled to meet the demands of platform engineering with a limited team amidst significant technological complexity. By leveraging AI, the company aims to ensure that small teams can maintain a sustainable pace of work and achieve better work-life balance, while still delivering robust R&D and CI/CD capabilities.
"If you leverage AI, you can actually make it more sustainable, you can find work-life balance, you could reduce toil, you could have a sustainable velocity."
Cisco Open-Sources 'Agency' Initiative to Standardize Internet of Agents
Cisco has open-sourced an initiative called "Agency" under the Linux Foundation, aiming to define the foundational infrastructure for interoperable agentic AI systems, often referred to as the "Internet of Agents." This project seeks to standardize critical components necessary for effective agent communication and collaboration, regardless of vendor or framework. Key areas of focus include establishing agent identities, discovery mechanisms, messaging protocols, deployment strategies, observability, and robust security frameworks to counter emerging threat vectors.
The vision behind Agency is to enable diverse AI agents from different organizations to seamlessly work together, fostering a collaborative ecosystem where agents can communicate and make decisions autonomously for specific tasks. By creating a common infrastructure, the initiative addresses the growing need for secure and standardized methods for deploying and managing AI agents, which are expected to interact extensively in future computational environments.
"Cisco did this collective call Agency and we open sourced it last year, donating it to Linux Foundation, so that's a very high level on that."
Outshift Developers Leverage AI Agents Across Multiple Workflows
Outshift developers interact with their AI platform engineering agentic system through various existing workflows, including an internal developer portal built on Backstage, WebEx, Jira tickets, and integrated coding environments like VS Code. This flexible access allows the AI system to assist with a wide range of tasks, from setting up CI pipelines and providing LLM access keys to handling general SRE and DevOps activities end-to-end.
The system integrates advanced functionalities such as knowledge base querying using graph RAG, which can extract insights from disparate documentation sources in minutes, a task that would typically take hours for a human expert. It also offers self-service capabilities that streamline form-filling for common requests and can generate complex Kubernetes configurations from natural language prompts, significantly reducing toil and improving developer experience by providing intuitive and immediate support.
"If you can get the information from segregated data sources into a graph database and letting agents do graph RAG on top of it is extremely powerful."
Successful AI Agent Deployment Requires Strong Platform Engineering Foundation
Effective implementation of AI agentic systems relies heavily on solid platform engineering and automation fundamentals, according to recent insights. Businesses are advised against focusing solely on AI metrics and should instead prioritize measuring overall business and team outcomes to truly assess the technology's value. Agentic AI should be viewed as an enhancement, or "icing on the top," that amplifies existing robust automation practices, rather than a standalone solution.
The Community AI Platform Engineering (CAPE) initiative, available at cape.io, promotes a community-driven approach to redefine this technological space. This open-source effort encourages shared innovation and best practices, emphasizing that foundational engineering discipline must precede and support AI integration. This ensures that AI agents contribute to consistent, repeatable, and genuinely impactful improvements, rather than becoming a costly distraction.
"My two cents here is don't measure AI, measure outcomes, because whatever business outcomes, team outcomes that you're after, measure those and actually evaluate whether AI is helping."
Observability and Evaluation Crucial for Probabilistic AI Agent Systems
Observability and evaluation are critical components for the successful development and deployment of probabilistic AI agentic systems. Projects are underway to extend OpenTelemetry, an industry standard for telemetry data, to specifically integrate agentic aspects, ensuring comprehensive monitoring. Additionally, a dedicated evaluation framework is being developed to rigorously assess the performance of AI agents.
These initiatives are vital for achieving accuracy, consistency, and repeatability in AI development, especially as agent behaviors can be unpredictable due to their probabilistic nature. By thoroughly observing and evaluating agent performance, developers can ensure that these systems operate reliably and meet desired standards, addressing a key challenge in bringing sophisticated AI applications to production.
"Evaluation is an absolutely crucial topic in order to get the accuracy and to make sure that you are making it consistent and repeatable."
Cloud-Native Complexity Strains Platform Engineering Teams
The exponential increase in complexity within the modern cloud-native stack poses significant challenges for running and maintaining production-grade applications. Unlike simpler software environments of a decade ago, today's ecosystems, dominated by microservices and diverse cloud technologies, often overwhelm platform engineering teams. While platform engineering aims to streamline operations, it frequently creates bottlenecks that can lead to developer frustration or team burnout if not managed effectively.
This inherent complexity can result in friction between development teams, who struggle with inadequate tooling, and platform engineers, who face cognitive overload trying to keep pace with the vast demands. Without competitive offerings and efficient management of this bottleneck, organizations risk slower releases and a decline in team morale. The situation underscores the urgent need for solutions that can simplify and sustain operations in these intricate environments.
"If your bottleneck is not able to offer a competitive offering, it's going to cause a lot of friction. It can lead to developers getting frustrated if your tooling is not up to scratch."
Visualizing AI Agent Task Planning Reveals Key Challenges
A visualization of a deep AI agent's task execution illustrates its complex planning process, involving multiple large language model (LLM) calls and the parallel invocation of several sub-agents to achieve a goal within 40 minutes. The system efficiently gathers data and presents results to the user. This demonstration, however, also highlights significant challenges inherent in probabilistic agentic systems, including issues with observability, evaluation, and continuous integration.
Successfully adopting these AI systems requires addressing complexities such as ensuring security governance against privilege escalation, managing the variability of agent behavior with model changes, and gaining customer confidence. Furthermore, a fundamental mindset shift is necessary within teams to effectively integrate and leverage AI. These challenges underscore the need for robust frameworks and practices to ensure the reliability and security of advanced AI deployments.
"You have like observability and evaluation challenges because these things are probabilistic in nature. Continuous integration is quite tricky because you change the model, you introduce an agent, the behavior can change."
Also mentioned in this video
- A fundamental paradigm shift is occurring in compute, moving from deterministic… (2:39)
- The spectrum of AI systems, differentiating between co-pilots, which require… (4:27)
- The speaker details various agentic systems implemented at Outshift for… (7:18)
- The high-level architecture of their software delivery platform, which includes… (10:00)
- The Agency framework includes verifiable identities for agentic systems,… (35:29)
- The open-sourced directory serves as a decentralized system for discovering… (37:00)
Summarised from Outshift by Cisco · 46:11. All credit belongs to the original creators. Streamed.News summarises publicly available video content.
