You text a friend: "running 10 min late, sorry." The reply arrives in under a second: "No problem at all! Thank you for letting me know. I hope your commute goes smoothly, and please do not worry about the delay. See you soon!"
Your friend has not been replaced by a robot. But for a moment, it feels like they have. Something is off. The reply was too fast, too long, too complete, and too eager.
That feeling has a name among people who build conversational agents: the uncanny valley of text. And if your agent talks to real people over SMS or chat apps, you need an opinion about it.
Here is mine: your agent should not pretend to be human, but it should respect how humans read. Those are two very different things.
Length: match the room
Text threads have a rhythm. A person who sends six words and gets back six paragraphs feels like they asked for the time and received a lecture.
A useful rule of thumb: mirror the other person's effort, then round down.
Agents default to length because language models default to length. Every sentence feels free to generate. But the reader pays for each one, on a small screen, probably while doing something else.
Bad:
> "Thank you so much for reaching out! I would be delighted to help you with rescheduling your appointment. Our available times next week include Tuesday at 2 PM, Wednesday at 10 AM, and Thursday at 4 PM. Please let me know which works best for you, and I will make sure to confirm it right away!"
Better:
> "Happy to move it. Tuesday 2, Wednesday 10, or Thursday 4?"
The second one is not less polite. It is less loud.
Timing: instant is a tell
Humans take time to reply. Not always, but often enough that the pattern is part of the language. A reply in 300 milliseconds says "I was already waiting for you," or more accurately, "I am a process."
Should you add artificial delay? It depends on what you are optimizing.
Add a short delay when:
Skip the delay when:
There is a line, though. A delay that exists purely to fool someone into thinking a person is typing is a small deception. The whitepaper's whole framing around identity and disclosure is that recipients should be able to know they are dealing with an agent. Pacing your replies so they are comfortable to read is courtesy. Pacing them so nobody notices you are software is a different thing.
Burst or single message?
Humans often send three short messages instead of one long one. It feels natural because it mirrors speech: a thought, a correction, a follow-up.
For agents, splitting a message into a burst has real tradeoffs:
| Approach | Good for | Risk |
|---|---|---|
| One message | Instructions, confirmations, anything the reader may save | Can feel like a wall of text |
| Short burst | Conversational flow, separating a question from context | Each message is a notification buzz |
| Burst with pauses | Casual, human-paced chat | Messages can arrive out of order if you are careless |
My rule: one idea per message, and never more than two or three in a burst. If you find yourself sending five, you wanted one message.
And never split anything the user might need to copy. Do not send a verification code, an address, or a link in the middle of a burst where it gets buried.
Typing dots: honest or theater?
Some channels show a typing indicator. It is a signal that says "something is coming," and it is genuinely useful when your agent needs a few seconds to think or look something up. A silent gap feels broken; a typing indicator feels alive.
Use it for what it means: work in progress. Do not use it as stagecraft, such as showing dots for eight seconds before sending a ten-character reply. Fake effort is still fake.
Also remember the indicator only exists on some channels. Plain SMS has no typing dots at all, so on SMS the only signals you have are timing and content.
When machine speed is a feature
Sometimes the best thing about your agent is that it is not human.
In these cases, speed is the service. Slowing down to seem human would be a worse product.
A quick checklist
The takeaway
The goal is not "text like a human." The goal is text like someone who respects the reader's time and attention. Humans are just the best-known example of that. An agent can be fast when speed helps, slow when pace helps, brief nearly always, and honest about what it is throughout.
Dial handles the channel plumbing and disclosure rules at the infrastructure layer. How your agent sounds is still your craft.