The Agentic Sweet Spot? The Human IS The Loop

In recent weeks, I’ve been hearing more and more chatter signaling disillusionment with agentic programming, as increasing numbers of once-upon-a-time very pro-agent folks walk back to taking much more control over their workflows. The available data suggests this was inevitable. While the horizon for agent autonomy has been roughly doubling every 6-7 months since 2019, according to METR, the horizon for reliable autonomy remains really, really short. Measuring AI Ability to Complete Long Tasks, METR, 2025 The upshot is that – if it needs to work – the human has to stay very much in the loop. In fact, when reliability really matters, the human has to be the loop, controlling every step in the process. While the “software factories” folks extol extending autonomous horizons as a higher level of AI use maturity, it’s becoming clear that the even higher levels dial back the autonomy to look for a sweet spot where software that’s reliable enough is being shipped earlier and at lower cost.…