Agents are here → AI’s shaken and stirred

From the shadowy corridors of business intelligence where precise decision making is pivotal, emerges the paradigm of the AI agent – a computer program capable of perceiving the environment and taking actions in relentless pursuit of goals and objectives. Think James Bond. Think Ethan Hunt. Except, this one’s not of flesh and blood, but of code and logic, a master of strategy and execution. Think more of Neo (The Matrix). When multiples of such independent but interactive agents are combined, each capable of perceiving the environment and taking actions, a multiagent system is birthed. Much like Hunt’s trusted IMF operatives, transferring outcomes seamlessly among the agents, each step pushing them closer to their objective, agentic AI defies the odds, accomplishing missions that others deem impossible. All, under the unyielding pressure of the clock. The AI agent is the digital embodiment of Hunt’s daring spirit – fearless and unstoppable. 

AI Agents - What exactly are they? 

Not always an AI model, an AI agent could even be a software program or a robot capable of sensing the environment and executing actions. In a multi-agent framework, the system understands a comprehensive objective, deconstructs it into a sequence of actionable steps and executes the sequence, transfers outputs from one stage to the next without any friction, culminating in a cohesive outcome. The system exhibits autonomy, as it independently strategizes and initiates actions based on its evaluation of the circumstances. An agent, or a group of them, can decide to engage with various interconnected systems and may even consult with a human if the situation demands it. 

How it works?

In multi-agent systems, complex plans are broken down into manageable sub-goals, which are then assigned to specialized "executor agents" adept at handling specific tasks. During a coding challenge, Andrew Ng's team evaluated the efficacy of an LLM on the HumanEval benchmark—commonly used to assess code functionality. Initially, GPT-3.5 alone achieved a 48% accuracy rate with mere prompting. However, when integrated into an "agent loop"—where agents employed strategies such as reflection (self-assessment and improvement by the LLM) and planning (developing and executing a multi-step approach)—GPT-3.5’s accuracy soared to 95%. The experiment demonstrated that distributing coding responsibilities among agents, each acting as a software engineer, product manager, or QA engineer, yielded optimal results. This was particularly effective when each agent concentrated on a singular task at a time.

IRL, Microsoft's AutoGen project showcases the superior performance of multi-agent systems over individual agents in managing complex tasks. It demonstrates the ability of collaborative agent systems to solve challenging problems that single-model or single-agent methods find difficult. 

How are agents different from AI or ML?

Traditional AI models provide output in a single pass; agents mimic human problem-solving by gathering data, drafting solutions, reflecting on weaknesses, and refining their approach. Rather than being based on rules, statistics, and thresholds, agents leverage deep and reinforcement learning through neural networks, allowing them to improve autonomously over time. While AI models function like advanced Q&A bots, agents act as accountable project managers, transforming AI into an active collaborator. NASA's Jet Propulsion Laboratory exemplifies this by using multi-agent systems to monitor clean room environments. These systems, which assess multiple contamination metrics collectively, offer a cost-effective solution compared to previous sensor technologies, highlighting their practicality and efficiency – albeit at lower sensitivities.

Now, what changes with agents?

As AI takes on more tasks, designing agent-optimized interfaces becomes crucial. Innovations like the SWE-agent – a system that facilitates LM agents to autonomously use computers to solve software engineering tasks – are outperforming RAG, solving 12.5% of issues versus RAG’s 3.8%. This research emphasizes that when agents manage work, interfaces must be tailored specifically for computers, not humans. AI-friendly APIs will offer significant competitive advantages by minimizing errors associated with human interfaces. Delegating tasks to agents presents challenges, such as tool-calling accuracy. For instance, with an 87% accuracy rate, completing a task requiring five sequential tool calls only achieves a 49.8% success rate. Furthermore, agents must "remember" information for extended tasks using real-world data, yet current LLMs often struggle with long-context tasks, relying heavily on synthetic data. 

Where are agents being used now?

First, projects are triaged based on risk levels: low-risk cases involve back-office functions, medium risks deal with internal data, and high-risk initiatives concern external users or sensitive data. Adopting such a well-defined stratification approach, insurance major Aflac adopts AI and yet ensures human oversight to prevent AI from operating entirely autonomously. Since 2018, NASA's Jet Propulsion Laboratory has utilized agents with Raspberry Pi computers and neural networks to maintain clean room standards, replacing costly sensors with a cooperative network that is both efficient and economical. Hughes Network Systems employs agentic AI to manage over a million remote nodes in its satellite networks. Initially, AI offered recommendations for human review, but now autonomously implements and reports changes by continuously analyzing network data for anomalies. Legendary Entertainment uses AI agents to detect movie leaks and piracy on social media and the dark web, leveraging an enterprise-level solution for automating business processes, although specific frameworks remain undisclosed.

What prevents the automation from going rogue and impacting the network?

Alarms. They monitor anomalies and corrective mechanisms, compare them to statistical norms and suspend actions if norms are exceeded. This prevents systems from behaving unexpectedly and ensures network stability. Since it’s Gen AI that powers the agents, leakage of creative content and assets through AI applications and the generation of infringing content are real threats. And so are attack vectors such as classical SQL code injections or database-stored procedure attack. 

And, so…

Promising solutions are on the horizon as new mitigation techniques emerge. Use of data loss prevention patterns to limit or exclude data types from being learned, inspection of instruction sets as modules learn and grow help prevent injections, hallucinations, and malicious code.

Alchemy in the archives: Vue.ai quickly adjusts to new document formats without extensive data collection or model retraining. Accurately processes complex financial documents, contracts, and forms, minimizing errors and ensuring compliance. Eliminates the need for costly data preparation and annotation. Handles evolving document requirements as business grows.

Shift happens; our AI orchestration platform responds attentively. Who benefits? Data teams (in unlocking cost-effective platform capabilities) ML teams (that help build and deploy a wide range of models) Business teams (to automate a variety of workflows). Pick Vue.ai to be your running mate. All your constituencies will love it. Dial in.