Chatbots and large language models can execute a seemingly countless number of tasks, from writing emails and reports to generating code and analyzing data. However, they still primarily act only in response to user prompts, and rely on their own predictive models to generate text. That's why they can be so good at some tasks, such as simulating human writing, and surprisingly bad at others, such as mathematical or logical reasoning.
But what if LLMs had the ability to reason, plan and use tools on their own, with limited human supervision? For example, what if an LLM could recognize that multiplying two numbers together is a math problem, and use a calculator to help it get the answer? That's the idea behind agentic AI, which experts believe is the next phase in the progression of AI.
Earlier this month, leading AI researchers from UC Berkeley and from top companies including OpenAI, Google, Amazon and Meta gathered on the Berkeley campus to discuss the current state and future development of agentic AI. Organized by the Berkeley Center for Responsible, Decentralized Intelligence, the Agentic AI Summit 2026 drew an estimated 5,000 in-person attendees and tens of thousands more online.
To learn more about agentic AI, UC Berkeley News interviewed Dawn Song, a professor of computer science and co-director of the center. In an email conversation, Song discusses the unique safety challenges posed by agentic AI, where the field is heading next and how events like the Agentic AI Summit and Berkeley's agentic AI massive open online course (MOOC) series are helping to make AI accessible beyond the ivory tower.
Berkeley News: What is agentic AI, and how does it differ from traditional large language models and chatbots? What new capabilities does it unlock?
Courtesy of Dawn Song
Dawn Song: Traditional large language models and chatbots are primarily designed to generate responses to user prompts. Agentic AI represents the next stage in that evolution. Instead of simply generating content, agentic systems can reason, plan, use external tools and take actions, execute multi-step tasks and adapt their behaviour as they interact with dynamic environments.
One way to think about it is the difference between asking someone for information and asking them to accomplish a goal on your behalf. An agent can gather information, use software tools, write and execute code, coordinate with other systems and continuously refine its strategy as new information becomes available.
This shift has the potential to fundamentally expand what AI systems can do. It opens exciting opportunities in areas such as scientific discovery, healthcare, software engineering, education and many other fields. At the same time, it raises important research questions around reliability, security and governance.
At Berkeley, we see these challenges not as obstacles, but as opportunities for foundational research that will shape the next generation of AI systems.
How will agentic AI change how we use these technologies?
I believe we're moving from AI as a tool that people consult to AI as a collaborator and capable partner that can help accomplish increasingly sophisticated tasks. Rather than assisting with one step at a time, agentic systems can coordinate multiple steps, make decisions within well-defined boundaries and adapt as circumstances evolve.
For individuals, this could mean AI assistants that help manage complex projects, support lifelong learning or assist with scientific and creative work. For organizations, it creates opportunities to rethink entire workflows across fields such as healthcare, scientific research, software engineering, finance, manufacturing and public services.
The opportunity is tremendous, but so is the responsibility. As AI systems become more autonomous, ensuring appropriate human oversight, transparency, accountability and alignment with human values becomes increasingly important. Building safe, secure, trustworthy AI systems will be just as important as building more capable ones, and that is an area where universities have a critical role to play.
What are some of these safety and security risks, and how can researchers ensure agentic AI systems remain reliable?
Agentic AI changes the security landscape because these systems are no longer limited to generating information; instead, they can take actions in the world. They may interact with software systems, access external tools, execute code or coordinate with other agents. As their capabilities expand, so does the potential impact of mistakes, unexpected behaviors or malicious manipulation.
For example, prompt injection attacks - in which LLMs are given inputs that effectively "trick" them into ignoring user instructions and instead perform malicious actions instructed by the attacker - can influence an agent. Errors may accumulate across long sequences of decisions, and agents may pursue objectives that differ from what users intended if they are not designed with sufficient safeguards.
Addressing these challenges requires advances across multiple dimensions: rigorous evaluation methodologies, secure-by-design system architectures, improved interpretability and stronger mechanisms for human oversight. At Berkeley, we view safety and security as fundamental research questions that should evolve alongside advances in AI capability, not after those capabilities have already been deployed.
Agentic AI Summit 2026
Since Berkeley launched its Agentic AI MOOC series in 2024, nearly 40,000 people have taken the course. Why is it important to ensure that AI education is accessible to a wide range of communities?
AI is advancing at an extraordinary pace, and it's important that our understanding of these technologies keeps pace as well, not only among AI researchers, but also among students, educators, policymakers, entrepreneurs and the broader public.
One of Berkeley's longstanding strengths is making frontier knowledge broadly accessible. The Agentic AI MOOC series reflects that mission by bringing cutting-edge research to learners around the world, regardless of where they are or what career stage they are in.
It's been incredibly encouraging to see nearly 40,000 learners worldwide engage with the material - from undergraduate students to industry practitioners and government leaders. The future of AI will not be determined solely by breakthroughs in research. It will be shaped by millions of people making decisions about how these systems are designed, deployed, governed and used. Expanding access to AI education is therefore one of the most important investments we can make - not only to accelerate innovation, but to ensure that AI is developed and applied wisely, responsibly and for the benefit of humanity.
How did the Agentic AI Summit 2026 reflect Berkeley's leadership in AI development?
Berkeley has a long tradition of advancing foundational computer science while also asking important questions about how technology should benefit society. I think the summit reflected both parts of that tradition.
Agentic AI has reached an important inflection point. The pace of technical progress has been remarkable, but many of the most significant challenges ahead involve evaluation, security, governance and responsible deployment. Addressing those challenges requires collaboration across academia, industry, government and the broader research community.
One of Berkeley's unique strengths is our ability to convene those communities. The summit was designed not simply to showcase new technologies, but to foster thoughtful discussions about where the field is headed and how we can guide its development responsibly. I hope it demonstrates the important role universities play, not only in advancing scientific breakthroughs, but also in helping shape the principles, talent and research ecosystem that will define the future of AI.
What are the biggest technical challenges facing agentic AI?
Several challenges stand out.
The first is reliability. Today's agentic systems can perform remarkably well on many tasks, but they still struggle with long-horizon reasoning, recovering gracefully from unexpected situations, and maintaining robust performance in complex real world environments.
Ultimately, I believe the next major breakthroughs in AI won't come from scaling models alone. They'll come from making AI systems more reliable, more trustworthy and better aligned with human goals.
Dawn Song
The second is evaluation. As AI systems become increasingly autonomous, we need much more rigorous ways to measure whether they are reliable, secure and trustworthy under realistic conditions in broad domains. Developing open benchmarks and scientifically grounded evaluation methodologies is an area where the academic community, and Berkeley in particular, can make especially important contributions. For example, at Berkeley, we have been exploring new evaluation frameworks, including platforms such as AgentBeats, enabling a new paradigm for standardized, reproducible agent evaluation.
Safety and security is another major priority. Agentic systems introduce fundamentally new vulnerabilities and failure modes that require new approaches to AI security.
Where do you see the field going next?
I'm excited about AI for cybersecurity. Increasingly capable agents can dramatically strengthen cyber defense by discovering vulnerabilities, verifying software, automatically generating patches and continuously monitoring complex systems. Given the dual use nature of AI, we also need to mitigate the risks of misuse of frontier AI by attackers.
I'm particularly excited about building AI systems that can generate software we can actually trust. Rather than current code generation where there's no security guarantee, we're beginning to combine AI with formal methods and program verification so that generated software can be accompanied by machine-checkable correctness and security guarantees. AI-assisted verifiable software development has the potential to fundamentally change how we build safety- and security-critical systems.
Ultimately, I believe the next major breakthroughs in AI won't come from scaling models alone. They'll come from making AI systems more reliable, more trustworthy and better aligned with human goals. Those are the foundational challenges that will determine how broadly society can benefit from increasingly capable AI.
