Order Matters

Try making a sandwich if you have not yet bought the bread. Try painting a wall you have not yet plastered smooth. Try publishing an article you have not yet written. These failures are obvious — yet when tasks become complex and numerous, the same logic quietly breaks down. You start a step before its prerequisites are ready, and the whole chain derails. In planning, the order of steps is not just important — it is often the entire challenge. Getting the steps right but in the wrong order produces the same result as not planning at all.

Dependencies: The Invisible Arrows

A dependency is a relationship between two tasks where one task cannot begin — or cannot be done correctly — until another task is complete. Dependencies are the invisible arrows that connect tasks in a plan. Dependencies come in several flavors. A finish-to-start dependency means Task B cannot begin until Task A is finished: you must bake the cake before you can frost it. A start-to-start dependency means Task B cannot begin until Task A has started: you cannot test the software until development has at least begun. A resource dependency means two tasks cannot run at the same time because they need the same thing: two students cannot use one microscope simultaneously. Identifying dependencies is one of the most important acts of planning — because every dependency you miss is a future collision waiting to happen.

Software engineers have thought deeply about dependencies for decades. When building a large program, hundreds of pieces of code may depend on each other. Engineers use a tool called a dependency graph — a diagram where each task or module is a node, and each dependency is a directed arrow from the prerequisite to the task that needs it. By looking at the graph, you can immediately see which tasks must come first and which can run simultaneously. AI planning systems use the same idea. A planning agent may build an internal dependency graph — sometimes called a task graph — before executing any action. This graph tells it: these tasks are blocked until their prerequisites are done, and these tasks can start right now.

Sequencing in AI Agents

When an AI agent receives a multi-step task, it needs to figure out not just what to do but in what order. Consider an agent asked to compile a research report on renewable energy. The raw list of tasks might include: search for recent studies, read and summarize each study, identify common themes, write the introduction, write the body sections, write the conclusion, check citations, format the document. These tasks have clear dependencies. Writing the body sections cannot happen before the studies have been summarized. Checking citations cannot happen before writing is done. Formatting the document must come last. But identifying themes and summarizing studies could potentially overlap if the agent has multiple tools running at once. A sequencing error — say, writing the conclusion before identifying the themes — produces a report that is incoherent: the conclusion refers to patterns that the body never established. The agent's output is complete-looking but wrong.