# What makes an AI agent's decisions reliable 2026-06-22 Usable inputs and explicit operating limits matter for agent decisions, alongside model uncertainty. The article asks where control and verification belong. In the audits I have run, including this site's own, one thing keeps surfacing. An agent that is instructed well, and given the right settings and checks, can take in data and make the decision the rules call for, consistently. The capability is real, and it is wider than most of the conversation around it. The limits are not always the model. They also sit in two places that are easy to overlook. ## A decision is only as good as its inputs The decision an agent reaches is bounded by the data that reaches the agent. In a clean datacenter that is invisible, so it gets ignored. Move the same agent to where the work actually happens and it becomes the whole problem. A link drops as a crane passes over it. A satellite hop can add roughly half a second to a full second, by my own rough estimate. On a transport that delivers in order, one lost packet can stall every packet queued behind it, and whether it does depends on the transport and the application. The agent waits on stale input while the moment it needed to act goes by. The agent did not get worse. Its inputs did. By my own reading of these audits, a large share of the reliability of an autonomous decision lives in the unglamorous layer below the model, where data either arrives in order and on time or it does not. A site or a system that wants an agent to act on live data has to earn that layer first. ## The settings decide what is allowed A correct decision is not an agent doing whatever it infers. It starts with an envelope that was defined for the agent, with its settings chosen ahead of time by a person who knew the stakes. Staying inside the envelope makes an action allowed, and an allowed action can still be wrong for the task. Draw the envelope loosely and a capable agent will still do something, just not the thing you wanted. Draw it well and the same agent can be left alone for longer on the task that envelope was built for, once its record on that specific task has been checked against a threshold set in advance. This post does not name a task, a threshold or a monitoring cadence, because none was measured here; a real deployment needs to name its own before making that call. This is the part that gets skipped when people picture autonomy. They imagine judgment appearing from nowhere. In practice the judgment is front-loaded into permissions and thresholds, and into an explicit list of what the agent may touch and what it may not. Good autonomy looks less like a clever model and more like a well-set boundary. ## The hardest case is where no one can step in The clearest test of all this is the environment where a person cannot be in the loop. Distance and latency, with help too far away to matter in the seconds that count. When the round trip to a human is longer than the decision can wait, the decision has to be made locally, under rules agreed in advance. The fields that operate in those conditions worked this out first, because they had no choice. They learned to package a human expert's judgment into something a machine could carry to the far end and apply without asking. That discipline used to look exotic. It is now the same thing any team needs before it lets an agent act on a system that matters. ## The point is not to remove the person Autonomy is not the absence of people. The strongest setups take an expert's judgment and place it where the work is, then let the machine handle the parts that have to be instant or exact. The person sees what the agent sees and acts through the same channel, and the agent extends their reach instead of standing in for them. This is why I have stopped describing my work as only agent-readiness. Reading a site is the first step, the precondition for everything after it. What an agent can actually do once the inputs are clean and the envelope is set, with a person kept where judgment belongs, is the rest of the distance. That is the work I am moving toward. For an agent-readiness audit, or a conversation about letting agents act on your systems safely, contact info@turva.dev. ## Frequently asked **What limits the reliability of an AI agent's decisions?** Not always the model. Two things sit below it. The data that reaches the agent, and the envelope of settings it is allowed to act inside. A decision is bounded by its inputs and by the settings. The settings decide which actions are allowed, and an allowed action can still be wrong for the task. **Why does the network layer decide whether an agent can act?** A link drops as a crane passes over it, a satellite hop can add roughly half a second to a full second by my own rough estimate, and on a transport that delivers in order one lost packet can stall every packet queued behind it. Whether it does depends on the transport and the application. The agent waits on stale input while the moment to act goes by. **What does good autonomy look like in practice?** A well-set boundary rather than a clever model. The judgment is front-loaded into permissions, thresholds and an explicit list of what the agent may touch. Draw the envelope loosely and a capable agent still does something, just not what you wanted. Corrected 2026-09-27. One sentence said a well-drawn envelope makes an agent one you can leave alone. Nothing in this post measures that, so the sentence now ties it to the agent's checked record. Corrected again 2026-09-27. A heading and two passages said a correct decision is the one the settings allowed. An allowed action can still be wrong for the task, so they now say the settings decide what is allowed and not what is right. Corrected 2026-09-28. A sentence said an agent can be left alone for longer once its record is checked, without naming what is checked or against what. It now ties that to the specific task's checked record against a threshold set in advance, and names that this post does not measure a task, a threshold or a cadence. Corrected 2026-09-28. Two more sentences stated an unmeasured comparison as fact. One now names a large share of decision reliability as my own reading of these audits rather than a measured share, and the other, which appears twice, now calls the satellite hop's added time a rough estimate of roughly half a second to a full second rather than a fixed figure. ## Related - [Letting agents act on data](/guides/letting-agents-act-on-data) - [AI agent use cases and their operating limits](/guides/ai-agent-use-cases) - [Agentic commerce readiness](/guides/agentic-commerce-readiness)