- Published on
Quoting Jon Udell
- Authors
- Name
- DeepHorus
Tech commentator Jon Udell has sparked a fresh debate about the role of artificial intelligence in software development, urging a shift in how the industry frames human-machine collaboration. In a recent post, Udell criticized the widely used phrase 'human in the loop,' arguing that it implicitly cedes authority to machines. Instead, he proposes a flipped narrative: 'human-agent in the loop,' where humans remain in control and invite AI agents to join their teams as collaborators rather than overlords.
Udell, a veteran developer and long-time advocate for open data and pragmatic tooling, has been at the forefront of integrating AI into everyday development workflows. His latest commentary arrives amid a surge of interest in agentic software development, where AI agents autonomously generate code, review pull requests, and perform tasks traditionally reserved for human developers. Udell worries that without careful framing, these agents could reduce humans to passive overseers.
'Itâs our loop, we work the same way we always have, now we recruit agents to join the team,' Udell wrote on June 28, 2026. He emphasized that an agent-assisted process need not be a 'black box that takes in prompts and emits features.' Rather, transparency and human oversight should remain central. His call echoes a broader sentiment among developers who fear losing agency to increasingly capable AI systems.
To illustrate his philosophy in action, Udell recently shared three concrete projects. On June 22, he ported the Moebius 0.2B image inpainting model to run entirely in a browser using Claude Code, demonstrating agent-guided but human-validated work. Two days earlier, he announced sqlite-utils 4.0rc1, adding migrations and nested transactionsâa tool that empowers developers with precise database control. On June 18, he launched Datasette Apps, enabling custom HTML applications inside the Datasette data exploration platform.
These projects underscore a practical approach where agents handle heavy lifting, but humans retain final authority. For instance, the Moebius port required iterative agent assistance, yet Udell supervised each step, ensuring the result was not a mysterious output but a transparent, debuggable artifact. This aligns with his broader vision: 'Letâs do agentic software development like that. Not as a loop weâve been excluded from, instead as one we invite agents into.'
The implications are significant. If Udellâs perspective gains traction, it could reshape how teams adopt AI toolsâprioritizing explainability, maintainability, and human judgment over raw automation. Critics might argue that ceding some control to agents is inevitable for efficiency, but Udell counters that the framing itself matters. By keeping humans 'in the driverâs seat,' teams can avoid the pitfalls of unreviewable AI-generated code.
Looking ahead, the industry faces a choice: embrace human-in-the-loop models that risk marginalizing people, or adopt human-agent-in-the-loop frameworks that uplift them. Udellâs recent projects serve as proof-of-concept for the latter. As AI agents become more capable, his call for transparent, human-led collaboration may become a blueprint for responsible innovation in software development.
