If you've noticed hackathon briefs increasingly asking for "agentic" solutions rather than just "AI-powered" ones, that's not just a terminology refresh - it reflects a real shift in what's being asked of teams.
Generative AI, plainly
Generative AI refers to models that produce new content - text, images, audio, code - based on an input prompt. You give it something, it generates something back, and that's the interaction: one request, one response. Most AI hackathon submissions from a few years ago fit this pattern - a chatbot wrapped around a language model, a tool that generates images from a description, and so on. Useful, but self-contained: the model doesn't take further action on its own after responding.
Agentic AI, plainly
Agentic AI describes systems where a model doesn't just respond once - it takes a sequence of actions to accomplish a goal, often using external tools along the way: looking something up, calling an API, checking whether the result actually solves the problem, and deciding what to do next based on that check. Often this is split across multiple specialised "agents" with defined roles rather than one model doing everything - one agent might handle research, another verification, another the final output - coordinated through a shared workflow.
EY Techathon 6.0's public brief is a concrete example of what this looks like in practice: teams build multi-agent workflows with defined roles, explicit tool permissions, and constraints on what each agent is allowed to do autonomously versus what requires human approval - not a single chatbot, but a coordinated system with guardrails.
Why this distinction actually matters for what you build
If a brief specifically asks for an agentic solution, judges are typically looking for evidence of genuine multi-step reasoning and tool use, not a single well-crafted prompt dressed up in agentic language. A team that builds a single-turn chatbot and calls it "agentic" without any actual multi-step, tool-using behavior is likely to be marked down once a judge asks how the system decides what to do next - because the honest answer, in that case, is that it doesn't.
What judges tend to actually check for
Whether the system can recover from a failed tool call rather than just breaking. Whether there's a clear boundary on what actions the system can take autonomously versus what needs human sign-off, especially for anything consequential (a real EY Techathon brief example: a loan-sales agent shouldn't be able to make an unauthorised credit decision on its own). Whether the team can explain, concretely, what happens at each step of the workflow rather than describing it only in vague, high-level terms.
You don't need a specific framework
Frameworks like LangGraph, AutoGen, or CrewAI get mentioned often in agentic-AI briefs as examples, but no specific one is usually required, and judges score the design and the outcome rather than the tool choice. A simpler, explicit state machine that your team fully understands and can defend clearly under questioning frequently scores better than an impressive-looking framework graph nobody on the team can actually explain when asked a direct question about it.
Where to see this in practice
The AI Hackathons 2026 guide covers the current programmes where agentic and GenAI themes actually show up in real briefs, including EY Techathon and Amazon's ML Challenge, with links to each one's specific eligibility and round structure.