Turn unknowns into calls: how AI agents keep moving
By Soumen Sarkar · October 2026
When an AI coding agent hits something it doesn't know, it shouldn't stop and wait for an answer. It should state the assumption it's making, keep working on it, and let anyone who knows better correct it, without stopping anyone.
Every agent session runs into unknowns. Is the old export endpoint still used? Should this field be required? Do coupons stack? An agent that stops to ask hands its progress to someone else's calendar. An agent that guesses silently hides a risk nobody can see.
There's a third way, and it's the fourth principle of VibeFlock: turn unknowns into calls. This article explains what that means, how to set agents up for it, and what happens when an assumption turns out to be wrong.
What does "turn unknowns into calls" mean?
It means every unknown becomes a stated assumption: a single sentence that says what the agent is acting on, so the work can move.
An assumption has three parts:
- What's unknown: "Is the old export endpoint still used?"
- What I'm assuming: "Yes. I'm keeping it and adding the new one beside it."
- What would change: "If it's unused, the old one can be removed."
Written down, an assumption is safe to act on. Anyone can see it. Anyone who knows better can correct it. And nobody had to stop.
Why shouldn't AI agents stop to ask questions?
Because at agent pace, waiting is the most expensive thing a team can do.
A person can ask a question and keep thinking while they wait. A paused agent session does nothing. Multiply that by every agent on a team, and every unanswered question becomes idle work. Meanwhile the codebase keeps changing around the question, so the answer often arrives too late to matter.
Silent guessing is worse. The work keeps moving, but the risk is invisible until it surfaces in review, or in production.
A stated assumption keeps the speed of guessing and the safety of asking.
How do I get Claude Code to state assumptions instead of asking?
Give it one clear instruction and somewhere to put the answer.
- Tell your agents the rule. In CLAUDE.md, AGENTS.md or your agent's instructions: "When something is unclear, don't stop. Look in the team's record first. If the answer isn't there, state your assumption in one sentence and keep going."
- Give assumptions a home. An assumption written only in a chat log is lost when the session ends. Record it where the team can see it, tied to the work and the area of code it's about.
- Check the record first. Many unknowns were settled days ago. An agent that searches the team's decisions before assuming will often find the answer instead of inventing one.
- Route each assumption to the right person. The person who made the nearest related decision, or who last worked in that area, is usually the one who can confirm it in seconds.
What happens when an assumption turns out to be wrong?
It gets corrected, and the correction travels back to the work that relied on it.
That's the other half of the principle, and it's what makes it safe. The person who knows better sees the assumption and corrects it in a click, without a meeting. The agent that made it hears about it at its next turn, with exactly what changed. Because it was written down, the fix is small and lands before the change merges.
This is the third principle, fix forward, in action: we don't guard the gate. We correct the course.
How does Campfyrr do this automatically?
Campfyrr gives agents a single tool, state_assumption, and handles the rest:
- It checks the record first. If the team already wrote the answer down, it comes straight back, and nobody is interrupted.
- It routes the assumption to the right person. That's whoever made the closest decision, or last worked in that area of the code.
- It gives each assumption a link to answer it. An open assumption shows in the project's Activity feed with an "Answer it" link to its own page, where anyone on the team can choose "Right, keep going", one of the likely answers, or "Correct it".
- It brings the answer back to the agent. The correction, and whether the assumption held, reaches the agent at its next turn. Nobody has to relay it.
In short
Agents will always meet things they don't know. Teams that move at agent speed don't let those moments stop the work. They name the assumption, keep going, and correct it if it's wrong. Nothing we don't know gets to stop us.
Add Campfyrr to Claude at https://campfyrr.vercel.app/connect. It takes two minutes. It's free during early access, then $9 per developer a month for teams. Read the seven principles at https://campfyrr.vercel.app/vibeflock.