This morning one of my AI agents sent an email to a leadership team I work with, telling them two Monday meetings should go ahead without me. It wrote: “Rajiv is on staycation Monday and Tuesday, so he won’t be at either meeting. He’ll read the transcripts afterwards, and he’s reachable by text.”
Third person. About me, not as me. And it read exactly right, which was awkward, because the rules I had written for my agents said the opposite: every message goes out in my first person, with the agent disclosed in a signature at the bottom.
The email was breaking my rules and improving on them at the same time. It took me the rest of the morning to understand why, and the answer turned out to be a job description.
Two agents, two jobs
I run two correspondence personas, and they are open-source projects you can look at: Ragbot and Ragenie. Until this morning I distinguished them by whose words they carried. Ragenie transmits words that are genuinely mine, composed or dictated or edited until I own them. Ragbot writes its own words under my direction, for routine traffic like scheduling, routing, and status.
The distinction that was missing is the one every executive office already has.
Ragenie is my chief of staff. A chief of staff is authorized to speak as the principal. When they draft, the draft goes out in my voice, under my name, and I own every word of it, which is why I review every word of it.
Ragbot is my executive assistant. An EA speaks in their own voice, for the office. “Rajiv is traveling this week; I’ve moved your 2:00.” A note like that speaks for the office without pretending the principal wrote it. Its third person tells you who is speaking, which is the whole reason the convention works.
So the rule I adopted, and then encoded: Ragbot speaks as itself. It says “I” meaning Ragbot, and it refers to me by name, in the third person. Ragenie speaks as me, because its entry condition is that the words already are mine.
The pronoun is the protocol
The obvious objection is that I already disclosed everything in signatures. Every agent-sent message ends with a line naming which persona produced it and how much I reviewed. Why does the grammar matter if the footer already tells the truth?
Because the footer is the most skippable part of a message. Signatures get cut from forwards, truncated in previews, dropped when someone quotes two lines of your email into a different thread. A third-person body cannot lose its disclosure that way. Every sentence carries it. No excerpt of “Rajiv is out Monday” can silently impersonate me, in any thread it gets pasted into, ever.
Once the grammar carries the disclosure, something better falls out: first person becomes a signal you can trust. If a message from my account says “I” and means me, I owned those words, full stop. My agents’ traffic can never dilute that, because their traffic never wears my “I.” The scarcer the real voice, the more it means when it shows up.
Mistakes land differently
Here is the operational reason this matters, not just the aesthetic one.
When an EA gets a detail wrong, the mistake is the assistant’s, and the principal corrects it without having personally said anything false. When a first-person message is wrong, the principal said something false, and no correction fully unrings that.
I still own everything my agents send. But there is a difference between owning a system’s mistake and having false words in my own mouth, and that difference is what determines how much I can safely delegate. The riskier an error sounds coming from me, the narrower the autonomous lane has to be. Move the routine traffic into the assistant’s own voice and the cost of a routine error drops, so the autonomous lane can widen. The voice model is not a style choice. It is what makes more delegation safe.
The hard limits did not move an inch, and this is worth saying plainly: the EA voice never carries my opinions, never makes commitments, never touches a sensitive relationship, never criticizes anyone. Those floors exist precisely so the lower-stakes lane stays lower-stakes.
What an assistant’s voice cannot carry
The model also took something away from Ragbot, and it was right to.
“Rajiv appreciates the quick turnaround” is worth less than “thank you for the quick turnaround.” Not slightly less. Categorically less, because gratitude routed through an intermediary announces that the grateful party could not be bothered to say it. The same is true of condolences, of congratulations, of anything whose entire value is that it came from a person.
So appreciation now routes to the chief of staff or to my own hands, always. My quick review of every appreciation message is the habit, not an inconvenience, because sincerity is the one thing I cannot delegate even in principle.
This gives me a routing test that beats any stakes calculation: would it sound right coming from an EA? A meeting move, yes. A thank-you, never. The question answers itself faster than any policy document could.
What the signatures say now
Each persona signs as itself, because a message should have one narrator from first word to last. Ragenie’s line is mine, spoken by me: “I wrote this with my Ragenie — synthesis writing: I personally write; AI assists.” The second half names the craft and links to it, and it earns its place by doing real work: most readers never click a signature link, so the clarification that synthesis writing is not machine writing has to travel in the line itself. Note the person of that clause. The craft’s site says “the writer writes, the AI assists,” which is right for a site; under my name the same idea has to be mine: “I personally write.” A slogan quoted in a signature is someone else’s voice in the one place that exists to mark voice. Authorship brackets the sentence at both ends. Ragbot’s lines are its own, and they vary with how much I reviewed. The one on standing-direction traffic reads:
🤖 I’m Rajiv’s Ragbot, sent under his standing direction — he reads every reply
That last clause is the oldest promise in the system, and it survived every redesign: whatever produced the message, the loop closes with me.
For the people whose titles I borrowed
I want to be direct about how this system should land with the people whose job titles I borrowed, because I borrowed them out of respect. The conventions work; I copied them.
If you are a human chief of staff, this is the machine that absorbs the mechanical load of your role so your judgment, your relationships, and your read of the room go further. If you are a human EA, same. The system runs on your profession’s conventions because your profession solved this problem long before software did. And if you are a manager who has neither, and most managers have neither, this is a working approximation of support you have been doing without, which is precisely why it is open source.
The mechanism is a public skill called synthesis-agent-correspondence in the synthesis ecosystem’s repository: the lanes, the voice rules, the signature patterns, and a persona registry where your own agents’ names go. My Ragbot and Ragenie are the worked example, not the product. Point the registry at your own names, write the standing limits you would give a new assistant on their first day, and the grammar does the rest.