Getting Started with AI Agents

Almost everyone in this room has used a chatbot. Far fewer have handed over a piece of work and had something go away, take twenty steps on its own and come back with it finished, which is a different skill and a much less forgiving one. That gap is where most of this year's disappointment actually sits: the tools are capable, and the work given to them is vaguely specified, unbounded and never checked. Haroon Choudery started an AI literacy nonprofit in 2017 to teach people whose jobs were about to change, co-founded a company building tooling for teams shipping generative AI, and now runs a practice that audits how an organisation really works before it builds anything for it. This session is for the people who have not started yet.

Who this is for

Anyone who uses a chatbot daily and has never built anything with one. Founders and operators being told to adopt agents without being told what that means, people who do the same task forty times a month, and anyone who tried once, got something unusable, and quietly put it down.

On the table

• What an agent is once the word is taken away: a model, the tools it is allowed to call, and somebody's decision about how far it may go unwatched
• Which work is worth handing over first, and why the answer is almost never the most interesting task you have
• What it costs when one is confidently wrong, and how you find out before your customer does

You'll leave with

• One task from your own week, specified tightly enough that something could actually take it
• A working sense of where to stop: what stays a draft for a human to approve, and what is safe to let run
• The questions to ask anything sold to you as an agent, and the answers that should worry you