Natural Language Is a Programming Interface. Context Is Part of the Input.
The latest message is only part of an agent's instruction. One quiet-hours example shows how context makes a brief request precise, and why keeping a superseded requirement can be worse than forgetting it.
“Move it beside the toggle.”
That can be a perfectly clear programming request. It can also be almost meaningless. The difference is what came before it.
When we work with an agent through conversation, the latest message doesn't have to repeat the whole task. It can change one part of an arrangement both sides already know. Judging it in isolation misses much of the input.
1. A short message can carry a specific change
Consider this hypothetical conversation about a quiet-hours settings screen:
User: Add quiet hours with an enable toggle, start and end times, and a summary of the saved period. Use the existing preference storage and the browser timezone. Leave notification scheduling alone.
User: Correction: use the account timezone, not the browser timezone.
Agent: I'll put the saved-period summary below the controls.
User: Move it beside the toggle.
The last message means: move the quiet-hours summary beside the enable toggle. It doesn't change how preferences are saved or expand the work into notification scheduling. The corrected account-timezone rule still applies.
Repeating all of that would make the message longer without changing its meaning. The conversation already supplies the details.
A new thread containing only “Move it beside the toggle” lacks those details. The agent needs the earlier task or the relevant artifact. If neither identifies the target, it needs a clarification rather than an invented screen to modify.
| What the agent has | What it can do with “Move it beside the toggle” |
|---|---|
| Earlier conversation | Move the quiet-hours summary and keep the account-timezone rule. |
| Latest message only | Ask which item and toggle the request refers to. |
These are the intended interpretations of this example, not results from a model test. The same words can be sufficient in one conversation and insufficient in another.
2. More history isn't always better context
The timezone correction is easy to overlook because the original request also sounds complete. An agent that remembers “browser timezone” but loses the correction has retained information and recovered the wrong requirement.
Context needs more than accumulation. The agent has to distinguish the current requirement from an abandoned option, an example, or its own earlier suggestion. Here, the later correction replaces one part of the brief; the layout follow-up replaces another. Everything else survives.
The same issue appears when work moves into a new conversation. A current task description saying “use the account timezone” is more useful than a long history whose competing statements have lost their order. Access to the repository may help recover the task, but it can't supply an agreement that was never recorded there.
3. Prompt length doesn't explain the whole interaction
Johannes Schneider's 2025 study of human–LLM conversations, using 2023 conversation data, found that user prompts tended to become shorter and simpler within conversations. The analysis was observational; its proposed explanation involving a shift in users' mental models remains conjectural. Shorter messages don't establish growing impatience or declining model intelligence.
Our quiet-hours example needs no diagnosis of the user's patience. The short request works because its missing details are already available. When they aren't, a longer shared history somewhere else doesn't help until the relevant information is recovered.
Natural-language programming lets us build on an established task instead of restating it every turn. What matters is whether the agent has the right meaning: the requested change, the rules that still apply, and the corrections that replaced earlier assumptions.