Navigating the Early Stages of Synchronous AI Agents

October, 2024

It has been nearly two years since generative AI (Gen AI) became democratized, enabling anyone—not just IT developers or data scientists—to interact with AI through simple natural language interfaces. But now, the question is, what if we could eliminate the need for these tech intermediaries and design AI systems that autonomously execute complex process workflows? This is where the industry is headed.

Currently, agents backed by large language models (LLMs) excel at handling singular tasks, such as creating a video, generating text, or analyzing content. However, executing a complete workflow still requires multiple exchanges, often with a human in the loop not only to prompt but to re-prompt when outputs fall short. This reliance on human oversight highlights a key limitation—AI, in its current form, is not yet fully autonomous.

The next frontier lies in building agents capable of self-learning and executing complex, synchronous tasks end-to-end from a single prompt. These agents would eliminate the need for constant human intervention and could dynamically adapt, reason, and refine outputs in real time. This evolution would mark a significant leap in AI capabilities, transforming how industries approach automation and process execution. The race to solve this challenge is where the industry stands today, pushing AI closer to true autonomy.

Defining Synchronous AI Agent

A synchronous AI agent is an advanced framework that streamlines complex workflows by breaking them down into smaller, manageable sub-tasks. Each sub-task is handled by specialized AI agents proficient in their respective domains.

For instance, during software development, one AI agent could focus on vulnerability testing, systematically scanning the code for security risks and weaknesses. Another agent might handle documentation, automatically generating and updating project documents to ensure that all team members have access to the latest information. Additionally, a third agent could handle code reviews, scrutinizing changes for consistency and adherence to best practices. At the same time, another might oversee continuous integration and deployment, ensuring that new code is thoroughly tested and seamlessly integrated into the existing software architecture.

This innovative framework not only boosts operational efficiency but also ensures that each critical component of software development is managed by an AI with specialized knowledge. By leveraging these targeted agents, organizations can streamline their development processes, minimize errors, and deliver higher-quality software products faster.

Screenshot 2024 10 27 at 7.48.08 PM 1030x588 - Navigating the Early Stages of Synchronous AI Agents

Figure 1: Components of a synchronous AI agent architecture

    • Centralized AI orchestration: A super AI governs the ecosystem, deploying specialized micro-AI agents (for example, content creators and fact-checkers) for distinct tasks. These agents collaborate autonomously or with human oversight when required. The super AI compiles and validates the information for final execution. It acts as a central architect, compiling and validating information, managing handoffs, and ensuring seamless coordination.
    • Project management: The system plans and organizes execution by identifying the sequence of sub-workflows and tasks within a broader process. For non-critical tasks, AI agents operate autonomously, while for high-risk processes, workflows are pre-coded and human-approved before AI agents act.
    • Seamless tool integration: AI agents operate within a predefined access framework, seamlessly interacting with external APIs and enterprise applications. This enables smooth and secure data flow and execution across tools, ensuring an integrated and frictionless ecosystem.
    • Autonomous action: Unlike traditional AI systems that provide recommendations and suggest the next best actions, AI agent autonomously takes action based on predefined guardrails, driving a higher level of operational independence while adhering to organizational risk parameters.
    • Contextual memory: AI agents retain and retrieve contextual information, ensuring continuity in tasks such as customer interactions, where understanding preferences, purchase history, or past engagements is critical for personalized marketing, sales, or product recommendations.
    • Dynamic adaptability: Unlike static, pre-programmed workflows, AI agents can reason and dynamically adapt to changing situations, adjusting their actions on-the-fly to meet real-time demands and evolving conditions.
    • Expert-level reasoning: AI agents possess reasoning abilities comparable to those of professionals such as doctors or financial analysts. This allows them to execute tasks with expert-level decision-making within strict organizational guardrails, ensuring decisions align with risk tolerance and business goals.
    • Multimodal perception: The AI agent processes diverse inputs, ranging from visual, auditory, and beyond, enabling it to perceive, interpret, and make decisions in complex environments with the depth of human understanding.
    • Continuous learning and improvement: AI agents engage in real-time, self-directed learning, continuously refining their performance by incorporating feedback and learning from prior outputs, ensuring adaptive growth and ongoing optimization.
    • Self-healing mechanism: In the event of system errors or breakdowns, AI agents autonomously diagnose and repair issues, maintaining uninterrupted operation and proactively ensuring system health without requiring human intervention.

Are Enterprises Ready for It?

The conversation around AI agents has gained momentum in the tech community throughout 2024. Leading voices such as Andrew Ng, founder and CEO of DeepLearning.AI, have actively encouraged enterprises to embrace this transformative wave of AI. One notable move was Atlanta-based startup Sem4.ai’s January 2024 acquisition of Robocorp , an automation vendor, to build intelligent agents capable of handling complex, end-to-end workflows. While this generated buzz, it is the developments in the past couple of months that have truly pushed AI agents into the spotlight.

In September 2024, major tech companies such as Salesforce, Oracle, ServiceNow, Microsoft, and AWS made groundbreaking announcements, launching AI agents designed to automate business tasks. Oracle introduced over 50 role-based AI agents for functions such as finance, supply chain, HR, sales, and marketing  while Salesforce unveiled a no-code platform,  enabling enterprises to create personalized AI agents with simple prompts. These moves are part of a broader rush to encourage clients to integrate AI agents into existing environments, automating tasks such as campaign management or IT service desk operations.

Case studies have emerged showcasing how companies such as Uber use AI agents to boost employee productivity. Notably, Klarna severed ties with Salesforce and Workday to fully leverage AI agents and plans to cut its workforce by half . However, while the tech sector is racing ahead, it is important to note that these companies have a high-risk tolerance, mature in-house AI capabilities, and a talent pool ready to build and deploy such systems.

For non-tech enterprises, such as those in retail and manufacturing, and for those in the regulated sectors, such as banking and healthcare, the road to adoption is far more cautious. These industries are unlikely to overhaul their stable, legacy systems for the promise of AI agents without thorough evaluation. While the technology is certainly exciting, it comes with inherent risks. Enterprises in these sectors need more than flashy product launches, they require solid ROI assessments, risk evaluations, and buy-in from stakeholders, including legal and compliance teams.

The interest, however, is undeniable. Enterprises want to understand how this “new kid on the block” will impact their workflows and competitive standing. But is this technology essential? Are competitors already adopting it? What impact will it have on existing infrastructure? Is it a passing fad or a long-term game-changer?

Currently, enterprises are in an educational phase—curious and excited but hesitant. Most are still experimenting with Gen AI proof of concepts (POCs) and pushing them toward production. While they see the potential, they are not ready to commit significant budgets. The path forward is promising but paved with challenges that need to be addressed before AI agent sees widespread, enterprise-level adoption.

Screenshot 2024 10 27 at 7.48.26 PM 1030x471 - Navigating the Early Stages of Synchronous AI Agents

Figure 2: Challenges faced by enterprises in synchronous AI agent adoption

    • Seamless collaboration and failure handling: Ensuring seamless coordination between AI agents is critical. A key challenge is managing what happens when an agent fails or stalls mid-process, potentially disrupting the workflow and leading to inefficiencies or errors.
    • AI hallucinations and contextual misinterpretation: While AI agents handle simple workflows effectively, they still struggle with complex tasks, especially nuanced understanding in natural languages, such as interpreting sarcasm in customer feedback or navigating ambiguous instructions. This limitation raises concerns about the trustworthiness of AI in high-stake scenarios.
    • Latency in time-sensitive applications: In industries where real-time decision-making is essential, such as financial trading or shop floor operations, any latency in agent responses can lead to costly errors or missed opportunities. Minimizing delays in time-sensitive processes is a critical hurdle.
    • Error compounding in sequential tasks: A 90% accuracy rate might seem impressive, but in agentic systems, even small probabilities of error compound as tasks become more sequential and interconnected. Achieving near-perfect reliability is essential for critical applications.
    • High computational costs: The computational resources required to run complex AI agent architectures significantly increase computing costs. Scaling these systems while maintaining efficiency and affordability remains a considerable challenge.
    • Workforce change management: As AI takes on more complex tasks, there is growing anxiety among the workforce about job displacement. Preparing employees to coexist with advanced AI, reskilling them for higher-value roles, and managing change will be crucial to successful AI adoption.

Strategic Steps for Early Adoption of Synchronous AI Agents

As organizations explore the potential of synchronous AI agents, it is important to approach implementation with a pragmatic mindset, starting small and scaling gradually. Here are three key recommendations for businesses at the early stages of adoption:

    • Start with unstructured processes: Before dismantling existing applications, such as Klarna’s bold move, organizations should focus on deploying synchronous AI agents in areas where processes are still evolving or not fully standardized. This allows teams to build POC models that demonstrate clear value without disrupting critical systems.
    • Don’t misinterpret “autonomous” capabilities: Despite the buzz surrounding synchronous AI agents, it is essential to recognize that true autonomy is still far off. LLMs are rapidly advancing, but they require significant human oversight and predefined guardrails to function effectively. These AI agents will need well-defined processes and approvals in place before they can execute tasks without human intervention.
    • Prioritize human-in-the-loop models: While synchronous AI agents promise to augment productivity, businesses should avoid planning for workforce reduction based on exaggerated claims in the market like those from companies such as Klarna or TikTok. For now, AI-driven synchronous agents should be viewed as tools to enhance human capabilities, with human oversight being critical to its success.

By Chandrika Dutt, Associate Research Director, Avasant, and Abhisekh Satapathy, Lead Analyst, Avasant