# What changes when the executor is AI? Author: Daniel Concepcion Published: 2026-08-14 Canonical: https://danielconcepcion.com/thinking/when-the-executor-is-ai/ > Useful progress between human decisions depends on explicit goals, feedback, authority, stopping rules and active human steering. I listen to AI podcasts, install local agents and automate research. But for me, nothing's better for learning a topic than testing it myself. Learning by doing. Along that journey, I wanted more useful work to happen between my decisions. That became the test for everything I built with AI. My first mistake was treating AI like a chat partner, meaning every single step stalled until I manually clicked 'approve'. So I moved to agents: - I gave bounded work to a named AI worker with authority, tools, context, permissions and feedback. - But activity still had no agreed finish. So I wrote the outcome and limits before work started. - Work then stopped after one step or circled the same problem. I defined success, failure, stopping and escalation. When the routes multiplied, I drew the paths and the evidence needed to move between them. We do the same with people. A job description goes nowhere without the systems, information, authority or budget required to act. AI needs a working environment too. The labels I found in the industry came later: - Agent: a worker with a defined role. - Goal: the outcome and its limits. - Loop: the feedback cycle for continuing, retrying or asking. - Graph: the map connecting workers, evidence and decisions. - Workflow: the predefined rules that move work through that map, like our processes, controls and escalation paths. From an executive seat, the pattern felt familiar to the way we work in our organizations today. We already assign ownership, decide authority, provide tools and information, review progress and handle exceptions. In this case, the 'employee' just happened to be an AI agent. Review gave me another familiar pattern. Recurrent issues survived until I used one model to do the work and another to evaluate it. The second caught issues the earlier review missed. Sound familiar? Peer review and separation of duties in our teams and organizations. I still set the acceptance bar and make the final decision. AI can follow a bad loop long after it stops producing useful work. At night, weak definitions get expensive. I added permissions, budgets, evidence requirements and stop conditions as software rules. Together, they form my deterministic safety net. At 2 a.m., I want hard rules steering the system while I'm asleep. I'm now measuring useful progress by the reduction in human interventions needed to keep work flowing. In short, how autonomy is progressing. Does that sound familiar from our own career journeys as we gain experience? As of today, to be honest, I am still testing which model fits each task and running the numbers on cost per accepted task to decide whether that autonomy is worth it. ## More Thinking - Newer: [Where AI lives](https://danielconcepcion.com/thinking/where-ai-lives/) - Older: [Black Hat USA 2026: the OpenAI–Hugging Face incident](https://danielconcepcion.com/thinking/openai-hugging-face-incident/)