The Intent Shift & The Autonomy Curve
Moving from syntax to intent, distinguishing traditional software from agents, decoding the 5 levels of autonomy, and 4 questions to validate agent opportunities.
Start with the shift in the product, not the novelty of the agent
The source paper describes a change from syntax to intent. People increasingly state the outcome they want while an AI system decides how to produce it. For a Product Manager, this changes more than the delivery process. It changes what the product must make visible, what can fail, and where human judgment belongs.
A conventional feature follows a path that the team has already encoded. An agent receives a goal, interprets context, chooses actions, uses tools, observes the result, and decides what to do next. That flexibility can create value when the path cannot be known in advance. It also creates a larger behavior surface that the product team must specify and evaluate.
The first decision on Day 1 is therefore not which model or framework to use. It is whether the user problem contains enough uncertainty and adaptive work to justify an agent at all.
Use one running example throughout this lesson: a product-feedback synthesis agent for a PM who receives interview notes, survey responses, and support tickets every week. The desired outcome is not “summarize text.” It is to identify recurring problems, attach source evidence, expose contradictions, and help the PM choose what to investigate. The agent may read approved feedback and draft themes, but it may not change roadmap priority or contact customers. This example is uncertain enough to require judgment, yet bounded enough to evaluate.
Read the autonomy curve as a product progression
The visual below illustrates how AI coding products moved from predicting syntax to executing goals. Each step transfers a different decision from the user to the system.
From Autocomplete to Autonomous Agents
Each generation preserved what came before while raising the ceiling on what one engineer could accomplish.
Coding Agents
(~2024-25)Human defines tasks and reviews diffs; agent runs iterative test-and-repair loop across files.
What stopping rules and token/time budgets prevent infinite execution loops?
Autocomplete predicts the next token while the user remains responsible for the plan. Inline generation can complete a larger unit, but the user still chooses where it belongs. Chat-based generation accepts a natural-language request and proposes an implementation. Coding agents go further by exploring a repository, editing several files, running tools, and correcting their own work. Autonomous agents can accept a bounded outcome and return later with a result.
For a PM, the important axis is not intelligence. It is delegated decision-making. As the product moves right on the curve, the team must answer more questions:
- What may the system decide without asking?
- Which tools and data can it access?
- How does the user see what it did?
- What evidence tells the system to continue, stop, or ask for help?
- How can an incorrect action be reversed?
The right position depends on the job and its stakes. A higher level of autonomy is not automatically a more valuable product.