AI agents in 2025: Expectations vs. reality (AI column)
When I studied the progress of AI Agents came across article About six months ago from IBM. In March 2025 IBM brought together four experts to discuss how real the AI agent revolution is. Among the experts,
- Maryam Ashoori, PhD: Director of Product Management, IBM® watsonx.ai™
- Marina Danilevsky: Senior Research Scientist, Language Technologies
- Vyoma Gajjar: AI Technical Solutions Architect
- Chris Hay: Distinguished Engineer
Their positions in this article looked something like this.
- Ashoori: Today’s “agents” are LLMs with basic planning and function calling. A true agent must reason and plan autonomously.
- danilewski (most skeptical)I still don't understand how this differs from regular orchestration. You just renamed orchestration agents because that's a buzzword.” Danilewski draws on programming experience - orchestration has been around for a long time. Ashoori shares current capabilities and theoretical potential.
- Hey. (most optimistic)The technological foundation is ready: Best, fastest, compact models -- Chain-of-thought training Advanced contextual windows -- Function calling Based on this: Hay analyzes the technical development of models 12-18 months and sees a qualitative leap in opportunity.
- Ghajar: “We’re seeing early glimpses, but complex solutions require breakthroughs in contextual thinking and edge cases.”
Disputes broke out among experts on the following topics: 1. Orchestra vs single agent Hey: There will be a pendulum movement - first a multi-agent system with an orchestrator, then a transition to "god-like" unified agents, then again to collaboration.
- Ashoori: This is an architectural solution, depending on the use case. You don't always need a meta orchestra. 2. Enterprise readiness Most organizations are not ready for agents. Interesting work will be in the exposition of the API of your enterprise systems.” Based on this: Understanding the current state of enterprise architectures and integration requirements. 3. Security challenges Ashoori: “What if an agent plugs into the dataset and deletes a bunch of sensitive records?” DANILEWSKY: Technology doesn't think. She can't be responsible. The scale of the risk is higher - technology can do more in less time unnoticed.” Gajar offers as solutions: strict testing in sandbox, rollback mechanisms, audit logs. 4. Human-in-the-loop vs replacement DANILEWSKY: Agents will augment people – “a person must be constantly present and make final decisions.” “There is a real risk that if implemented incorrectly, humans will complement AI rather than the other way around.”
What are the forecasts of experts based Technical Analysis: Hay looks at progress in modeling capabilities
- Practical experience: Danilewski builds on experience with current systems Enterprise research: Ashoori relies on research on developer behavior Architectural Understanding: Gajar Analyzes Requirements for Production System n
Conclusions for engineers Current “agents” are upgraded LLM with function calling. True autonomy is still far away, but the basic cases are already working. Value will be given to those who organize their data for agent workflow. The problem is not with models, but with enterprise-readiness. Transparency, traceability, rollback mechanisms are not options, but must-have. One uncontrolled agent can delete critical data. “Don’t be a hammer looking for nails” – first determine the ROI, then deploy the agents.
The main prediction of the authors is that 2025 It will be a year of experimentation with agents, but a revolution is to be expected later. Prepare the ground now, but keep expectations realistic. So far, it seems to be coming true:)
#AI #Software #Architecture #Agents #Leadership #ML #SystemDesign