Liveness in the Age of Agents: Are We Back to Compile-and-Run Cycles?

An experienced programmer who works happily motivated within a live programming environment can yield high-quality software within reasonable time. A novice can also perform well because live programming tools foster short feedback loops, which help translate thoughts into executable program code and thus the desired observable system behavior. In the age of agents, now, one might think that actual programming skills take a step back to let visions shine: express your goal in natural language and let the agents make even larger, coherent edits to the system. However, agents will take their time to think, and programmers require skills to make informed decisions about an agent’s generated results. How “live” is that? How “short” does this feedback loop feel? In this paper, we take a look at Tanimoto’s liveness levels 5 and 6 in combination with Norman’s interaction “gulfs” to discuss the impact agents have on the overall programming experience. We realize that the agent support is far from predictive as prompts take noticeable effort to clarify (or derive) a programmer’s current goals. We also conclude that while the interaction cycles might feel quite productive at times, the experienced “Gulf of Evaluation” can be too much for novices if the results are too sophisticated. For the (near) future, we believe that the programming experience around agents can further be tailored to improve both prompting and reviewing so that this very promising technology does not take us back to the times where lengthy compile-and-run cycles disrupted our thoughts, flow, and motivation.