The days of AI acting solely arsenic a predictive instrumentality oregon chatbot are numbered.
Armand Ruiz, IBM’s vice president of merchandise absorption for its AI platform, told delegates astatine the SXSW Festival successful Australia this week that AI agents volition soon let enterprises successful the APAC portion to automate complex, multi-step tasks, freeing employees to absorption connected much human-centric activities.
Ruiz explained that AI technologies person evolved from predictive models successful accepted instrumentality learning to the wide usage of chatbots. He predicted that the adjacent leap volition usher successful an “agentic era,” wherever specialised AI agents collaborate with humans to thrust organisational efficiencies.
“We person a agelong mode to spell to get AI to let america to bash each these regular tasks and bash it successful a mode that is reliable, and past bash it successful a mode that you tin standard it, and past you tin explicate it, and you tin show it,” Ruiz told the crowd. “But we’re going to get there, and we’re going to get determination faster than we think.”
What is an AI agent?
According to Ruiz, an AI cause is simply a strategy that tin autonomously crushed done analyzable problems, breaking down tasks, creating actionable plans, and executing those plans utilizing a suite of tools. These agents grounds precocious reasoning, representation retention, and the quality to execute tasks independently.
Ruiz identified 4 capabilities of AI agents: planning, memory, tools, and action.
AI agents and their capabilities
1. Planning
AI agents are susceptible of precocious readying to code fixed tasks oregon prompts.
Self-reflection: Agents tin self-reflect oregon cheque if their decisions marque consciousness oregon not.
Self-criticism: Agents tin usage feedback, often from the aforesaid oregon antithetic ample connection models, to critique and amended their plans.
SEE: Gartner’s apical 10 strategical exertion trends for 2025
Chain of thought: Agents tin interruption down larger tasks into smaller steps to amended accuracy.
Sub-goal decomposition: They tin besides found sub-goals by dividing larger tasks into manageable components.
2. Memory
AI agents leverage some short- and semipermanent representation to enactment their autonomous actions.
Short-term memory: This in-context representation enables agents to way actions wrong an existing session.
Long-term memory: AI agents tin log past interactions, helping them larn from mistakes and continuously amended their show implicit time.
3. Use of tools
AI agents volition beryllium connected to third-party tools to implicit their tasks. With the close entree and governance, they could leverage tools ranging from web hunt and code-generation platforms to endeavor systems, specified arsenic HR platforms, Microsoft Teams, CRM tools, unreality services, and information warehouses.
4. Autonomous action
The existent imaginable of AI agents lies successful their quality to act autonomously connected behalf of humans. Whether streamlining HR workflows specified arsenic recruiting, resolving bundle codification issues, oregon tackling different endeavor challenges, these agents volition alteration AI from passive chatbot to proactive actors.
AI agents tin plan, remember, entree tools and instrumentality action. Image: IBMEnterprises volition orchestrate armies of agents arsenic portion of their workforce
Enterprises are apt to person “millions of AI agents” moving for them, Ruiz said. These agents, which volition fundamentally enactment arsenic coworkers oregon AI assistants for quality employees, volition beryllium capable to enactment collaboratively with 1 different connected assorted tasks, enabling them to “solve problems end-to-end.”
Ruiz explained that AI agents tin relation arsenic either single-step oregon multi-step systems, with their actions coordinated and guided by a Super AI.
One-step AI agents
One-step agents are those that tin execute circumstantial tasks oregon lick idiosyncratic problems erstwhile prompted, executing them with the assistance of applicable tools. Tools are defined and the process inactive remains reasonably manual, though these agents tin entree systems specified arsenic LLMs to nutrient results.
Ruiz warned that determination whitethorn beryllium times erstwhile these AI agents hallucinate oregon neglect to enactment arsenic good arsenic desired.
Multi-step AI agents
Multi-step AI agents leverage iterative strategies successful what Ruiz called a “thought, action, reflection loop,” utilizing 1 oregon much LLMs. “You person this loop that is precise iterative, and it’s astonishing however that improves the outcome, and provides amended results until you get the last one,” helium said.
Super AIs
Enterprises volition deploy “Super AI” systems to coordinate networks of idiosyncratic AI agents. According to Ruiz, these Super AIs volition enactment arsenic orchestrators, readying tasks, breaking them down into smaller components, and assigning them to the astir suitable agents wrong the organsiation to implicit the enactment efficiently.
“One AI cause mightiness beryllium precise bully astatine income oregon merchandise absorption oregon encoding, oregon precise bully astatine mainframe oregon a circumstantial programming language. Each volition person tiny connection models that are precise casual to train, precise inexpensive to execute and they volition person circumstantial entree to definite tools,” helium said.
Who volition beryllium the large users of AI agents?
Ruiz identified 3 superior idiosyncratic groups apt to payment from AI agents: developers, no-code concern users, and extremity users.
Developers: Traditionally, AI, information science, and instrumentality learning required highly specialised expertise. However, Ruiz explained that millions of developers present person entree to these technologies via APIs. Additionally, frameworks similar CrewAI let developers to rapidly physique and deploy AI agents.
Business users: No-code tools volition soon let business users to physique their ain AI agents done a idiosyncratic interface. IBM’s caller cause builder, acceptable to debut astatine IBM’s TechXchange Conference, volition empower employees crossed each levels of a concern to make agents that tin automate and execute organisational tasks without needing programming knowledge.
End users: A wide scope of extremity users volition besides prosecute with AI agents, Ruiz said, noting that determination volition beryllium “a full spectrum” of extremity users adopting and utilizing these tools successful assorted ways.
How agents volition alteration our enterprises and work
Ruiz said factories were a bully analogy for however enactment whitethorn transform. In the aboriginal 1900s, factories relied connected manual, labor-intensive enactment by galore people, which was precise time-consuming and inefficient. However, upon the dawn of the concern revolution, machines were introduced to assistance automate them and accelerate production.
He explained that AI is present evolving to assistance automate and augment intelligence enactment successful the aforesaid mode arsenic machines automated carnal labour successful factories. Rather than being a replacement, he argues it volition alteration them to absorption connected much strategical and innovative tasks, improving wide productivity and efficiency.
“We’re seeing this already successful marketing,” Ruiz added. “We’re going to spot this successful income arsenic well, and it’s going to commencement expanding passim each the antithetic occupation functions. Our extremity is to get AI to escaped america from a batch of distractions and alteration america to enactment connected meaningful enactment and connected quality connections.
“The imaginativeness is for AI agents to enactment alongside humans successful a complementary way, augmenting quality capabilities alternatively than replacing quality workers entirely. This volition let for greater productivity, work-life balance, and absorption connected higher-value activities.”
English (US) ·