Essay 01 · Agentic systems
A Bet on Learning by Demonstration
Could this be the next step for agentic systems?
The most useful agent may not be one that knows every workflow in advance. It may be one that can learn the way we teach a new colleague.
Today’s agents can perform increasingly complex tasks, but they remain far less flexible than humans. Their underlying capabilities may be general, but their deployment is not: each new workflow still requires custom instructions, integrations and evaluations.
My bet is that the next step for agentic systems is learning by demonstration.
A user performs a task a few times while the agent observes. The agent infers the objective, identifies the important decisions and attempts the workflow itself. Feedback on the result helps it refine the process and generalize it to related tasks.
The hard part is not copying clicks. A demonstration shows what someone did, but not necessarily why. The agent must distinguish essential decisions from personal habits, recover missing context and infer what constitutes a correct outcome. Learning by demonstration must therefore be paired with verification: observe, attempt, evaluate, clarify and improve.
If this becomes reliable, agent deployment changes from an integration project into a teaching process. Companies no longer need to manually program the long tail of internal workflows. Domain experts become teachers: their demonstrations provide an initial approach, while real operating outcomes provide the feedback required to improve it.
I do not think the most useful agent will be one that knows every workflow in advance. It will be one that can be taught the way we teach a new colleague: show it the work, explain what matters, correct its mistakes and let it improve.