Intent Disambiguation in AI Agents: When Commands Are Ambiguous
Intent Disambiguation in AI Agents: When Commands Are Ambiguous
Someone told their AI agent to "walk the dog" and it started researching dog-walking business registrations. The agent interpreted a personal task as a business opportunity. Funny in a tweet. Terrifying if the agent had spending authority.
Intent disambiguation - figuring out what the user actually means - is one of the hardest problems in agent design. And it is one that gets worse, not better, as agents become more capable.
Why More Capability Makes This Harder
A simple agent that can only set timers and send messages has limited ways to misinterpret "walk the dog." A capable agent with browser control, file access, and the ability to make purchases has hundreds of possible interpretations. More tools mean more ways to be wrong.
The problem compounds with context. "Send that to Sarah" makes sense if the agent knows which document you were just editing and which Sarah you mean. Without that context, it might email the wrong file to the wrong person.
Three Disambiguation Strategies
Ask before acting. When the command is ambiguous, the agent should clarify. "Did you mean add 'walk the dog' to your to-do list, or are you looking for something else?" This is the safest approach but breaks flow if overused.
Use context to narrow intent. The time of day, recent actions, active applications, and your typical patterns all provide signal. "Walk the dog" at 7 AM while your calendar is open probably means "add to today's tasks." The same phrase while researching side hustles might genuinely mean "explore this business idea."
Rank by reversibility. When the agent must guess, prefer the interpretation with the lowest cost of being wrong. Adding a to-do item is easily undone. Registering a business domain is not. Default to the safer interpretation and let the user escalate if needed.
The Confirmation Sweet Spot
Over-confirming kills the value of an agent. If every command triggers a clarification question, you might as well do it yourself. Under-confirming leads to expensive mistakes.
The sweet spot is confirming only when the blast radius of a wrong interpretation is high. Reading a file based on an ambiguous command - just do it. Sending an email based on an ambiguous command - ask first. The agent should internally estimate the cost of being wrong and scale its confirmation behavior accordingly.
Good agents do not just execute commands. They understand them.
Fazm is an open source macOS AI agent. Open source on GitHub.