Module 1 · Foundations of Agentic AI · scripted
Every Response Is Part of a Future Prompt
Every response is part of a future prompt
When an LLM responds, its answer isn't merely an output. It becomes part of the conversation the model will read on its next turn. In that sense, every response is part of a future prompt. The model is continually helping construct the context that will shape what it does next.
Watch it happen. Pay attention to where the word Fjordlight comes from:
I need a name for a research project about coordinating autonomous ships in Norwegian coastal waters.
A few directions: Fjordlight (navigation and guidance connotations), Konvoi (cooperative movement), or Leia (the old Norse word for "to lead" — also a sailing route). Fjordlight is probably the most memorable.
Write a one-sentence project summary.
Fjordlight develops coordination methods that let autonomous vessels negotiate shared passage through Norwegian coastal waters safely, without centralized control.
The loop
So the model is doing two things at once: it is responding to the current conversation, and it is writing part of the future conversation — the one it will be given next time. The recursion is not in the model itself. Its weights don't change during a conversation. The recursion is in the conversation: the model participates in constructing the very context that will govern its future behavior.
Every answer changes the next question
The effect runs through the other participant too. What the model says shapes what you say next — so its output steers its own future input even when you write the words:
Help me plan a day around Molde.
Happily. First — are you drawn more to a mountain day (the Romsdalseggen ridge is nearby) or a coastal day (the Atlantic Ocean Road)?
Mountains, definitely.
Then here's the shape of the day: bus to the Romsdalseggen trailhead at 08:15, the classic ridge traverse (6–8 hours), and the evening back in Molde with the fjord view from Varden.
The formal view: a recursive computation over accumulating state
There is a compact way to write down what these examples show. An LLM conversation is a recursive computation over an accumulating state:
The recursion
Map it back to Example 1, turn by turn:
The recursion, unrolled on Example 1
The interesting part is the last term: the LLM computes an output which is then incorporated into the state on which the LLM will subsequently operate. Its previous computations become inputs to its future computations. That is all the formalism says — it is Example 1, written in symbols.
Conversational self-conditioning
Notice what this is not. The model is not self-modifying — no weights change, nothing is learned. The model is self-conditioning: its outputs alter the context that conditions its future outputs.
The progression
Early outputs have downstream effects
Self-conditioning has a consequence worth staring at: whatever the model writes early — a framing, an assumption, a term, a plan, an error — becomes part of the evidence it sees on every later turn.
How should I get from Molde to Ålesund for a morning meeting?
Driving is simplest: leave by 06:30, take the Vestnes ferry with your car, and you'll be in Ålesund by 09:00 with parking near the center.
What time would I need to get up?
To have the car loaded and make the 06:50 ferry, you'd want to be up by 05:45 and out the door by 06:20.
And if the weather is bad?
Allow extra margin for the drive to the ferry quay — leave by 06:00, and keep the car in the queue early, as crossings can hold in high wind.
- Which exact words introduced the car?
- What would the conversation look like if turn 2 had asked "car or bus?"
- What's the cheapest possible fix right now, at the end of this conversation?
It cuts the other way too — a good early structure becomes scaffolding that later turns keep building on:
Help me compare two PhD case-study sites. Give me your criteria first.
Four criteria: data access (can we instrument the operation?), variance (does interesting variation occur often?), partner commitment (will they stay for three years?), and generalizability (does it transfer beyond this site?).
Site A is a fish-processing plant that offered full sensor access.
Against the criteria: data access — excellent, sensor-level. Variance — promising, seasonal and catch-driven. Partner commitment and generalizability — still unknown; ask about staffing stability and how typical their process is.
LLMs don't just generate answers. They generate their future context.
Open a recent LLM conversation of yours and find one thing the model introduced — a term, an assumption, a structure — that later turns kept building on. Was it a car, or was it a scaffold?