For founders drowning in manual prompts, Claire Vo offers a sharp rebuke: “If your agent isn't able to prompt itself through an automation, what are you even doing?” This isn't just about handing off simple tasks; it’s about a deeper shift where AI agents become truly autonomous, prompting their own next steps.
The core insight is straightforward: move beyond one-off interactions. An AI agent doesn't need you to babysit every decision. Instead, you design "loops" that allow the agent to instruct itself, execute work, and even validate its own outcomes. Claire Vo calls this out directly: "when we're talking about loops, we're talking about automated prompting of an agent." This moves your AI from a reactive tool to a proactive, self-driving worker.
The Method: Designing Self-Prompting AI Agent Loops
Forget the idea that an AI agent needs constant human input. The goal here is to set up systems where the AI tells itself what to do next. Vo details two primary types of these self-prompting loops:
1. Scheduled Loops: These are your AI's version of a cron job or a webhook. Think "heartbeats, crons, hooks." They run at set times or trigger on specific events, giving the agent a predefined instruction. For example, you could set an agent to summarize competitor news every morning at 7 AM, or to flag new customer support tickets as they arrive. The agent doesn't wait for your command; it acts on a schedule.
2. Goal-Driven Loops: This is where the real power kicks in. Claire Vo defines a goal as “a type of loop that sets an outcome and runs an agent against that outcome until the outcome can be measured and validated or the agent is blocked.” Here, you give the agent a desired end-state, and it iteratively prompts itself, taking actions until that outcome is met. It might generate several drafts, run an internal check against criteria, revise, and repeat until it hits the target. Imagine an agent tasked with drafting five unique ad copy variations, then using a sentiment analysis model to rate them, revising any below a certain score, and stopping only when it has three high-scoring options.
This framework flips the script. Instead of you prompting, the agent is constantly prompting itself based on its current state and its overarching mission. It's not about complex AGI; it's about engineering simple, defined automata that keep working towards a measurable result.
Where This Breaks Down
While powerful, these autonomous loops come with serious caveats. The biggest one Vo highlights is token cost. An agent endlessly prompting itself, even efficiently, can burn through API credits at an alarming rate. Without tight guardrails, a goal-driven loop could become a runaway expense, rapidly draining your budget for minimal gain. You need a clear sense of when an agent is actually blocked, not just spinning its wheels.
Another pitfall is precision in prompt writing and outcome definition. If your initial prompt is vague or your "measurable outcome" is fuzzy, the agent will loop on bad logic. Claire Vo stresses the need for "precise prompt writing." An agent striving for an ill-defined "good marketing email" will produce inconsistent results indefinitely. The success of the loop depends entirely on how clearly you define the task, the iterative steps, and the validation criteria.
What to Do With This
Pick one repetitive, human-prompted task in your startup that has a clear success metric. Maybe it’s drafting social media posts, analyzing simple data, or generating customer replies. Now, instead of thinking about how you would prompt it, design a simple loop where the agent prompts itself.
Define the start: "Generate 3 tweet ideas for our new feature X." Define the iterative step: "If a tweet idea includes a forbidden word (e.g., 'revolutionary'), re-prompt to remove it." Define the measurable outcome: "Stop when 3 unique tweet ideas are generated, each under 280 characters and free of forbidden words." Ship this small loop, monitor its token usage like a hawk, and refine your definitions. Start small, but start automating.