Subgoals and Milestones

Running a marathon is 26.2 miles. No experienced runner thinks about 26.2 miles as a single thing to conquer. They break the race into segments: reach the 5-mile mark feeling controlled, hit the halfway point with energy in reserve, push through the 20-mile wall, sprint the final stretch. Each of those targets is a subgoal — an intermediate destination that keeps the runner oriented, motivated, and on pace when the finish line is nowhere in sight.

What a Subgoal Actually Is

A subgoal is a condition that must be true partway through a plan. It is not the final destination, but it is a real achievement that brings you measurably closer. Think of subgoals as stepping stones across a river — each stone is a firm place to stand before taking the next step. Subgoals serve four purposes. They make progress visible: you can check off a subgoal and know the plan is on track. They divide a long uncertain journey into shorter certain ones: you do not know what the full path to the final goal looks like, but you know exactly what you need to do to reach the next subgoal. They allow course-correction: if you reach a subgoal and something has gone wrong, you catch it early rather than at the very end. And they give AI agents a way to verify their own progress, since an agent can check whether a subgoal is complete before moving on.

A milestone is a special kind of subgoal: a clearly visible checkpoint that marks the end of a major phase of work. Where subgoals can be tiny and numerous, milestones are fewer and more significant. In a software project, milestones might be: requirements finalized, prototype running, testing complete, product shipped. Each milestone represents a phase boundary — the whole team can see it, celebrate it, and use it to decide whether the plan is still realistic.

How AI Agents Use Subgoals

When an AI agent plans a complex task, subgoals act as internal checkpoints. Consider an agent tasked with booking a business trip. A naive agent might immediately try to book a flight — but without first confirming dates, it might book the wrong day. A planning agent sets up subgoals: confirm dates, confirm destination, identify budget, find flight options, find hotel options, compare total costs, book both. At each subgoal, the agent can pause and verify: is this condition actually met? Did I confirm the dates? If yes, move on. If the airline website returns an error, the date-confirmed subgoal is not yet satisfied, and the agent knows to retry rather than blindly proceeding. This verification loop — check whether the subgoal is met, proceed if yes, retry or replan if no — is what separates a robust planning agent from one that barrels forward regardless of what is actually happening.

Designing Good Subgoals

Not every subgoal is equally useful. A good subgoal has three properties. It is verifiable — you can check clearly whether it is met or not. It is necessary — the main goal cannot be reached without it. And it is proportionate — it covers a meaningful chunk of work without trying to cover too much. A bad subgoal is vague: make some progress on the report. How do you check that? A good subgoal is concrete: complete the introduction section and have it reviewed by a partner. That is verifiable, necessary, and proportionate. For AI agents, vague subgoals are especially dangerous. An agent cannot reliably check whether it has made some progress. It can check whether a specific file was saved, a specific email was sent, or a specific search returned results. Concrete subgoals turn ambiguous intent into checkable conditions.