Why do I run two AI assistants with separate jobs?
Tao is the AI assistant I use as a head coach. Thy is my main executive assistant. I run these two AI assistants with separate domains on purpose, so each draws on sources that fit its job and keeps its own memory. Both run as Claude Code sessions that I talk to in chat channels, and each has a written charter that says what it owns and when it must stop and ask. This page covers two of the assistants I work with.
The alternative is one assistant connected to everything, and I did not build that. Every extra source is one more place for stale data to enter, and I rate fresh sources above broad ones. With two narrow assistants, a wrong answer points to a short list of sources to check.
What does each assistant actually do?
Tao looks after body, mind and career as one whole. Before any answer about health or training, Tao reads a living state file and is told that the file may be stale and has to be checked against the source systems. Advice is expected to cite the data behind it, whether that is a line in the state file, a training log or a reading. Medical questions go to a clinician.
Thy works the other side of the week: commitments, travel, follow-ups and the founder-alignment thread. Thy reads the channel thread before answering anything that assumes earlier context, and never answers from memory about a date, a person, a flight or a promise. The charter measures Thy by what is confirmed versus what is still open, always with the source. Personal finance and security material stays in the channel it came from.
The two charters share a shape. Each one names an objective, the sources the assistant should read, the judgment it is trusted with, the actions it may take alone and the cases where it escalates. Only the content differs.
Why do I avoid the phrase full context?
No assistant has it. Pretending otherwise is how stale information ends up driving a decision.
A real case from September shows the mechanism. A usage table that Thy posts for me showed one of my Claude accounts at 93 percent for 12 hours. The real figure was between 57 and 73 percent. A cache file was only rewritten on one code path, so it outlived the login that wrote it, and it got labelled with whichever account was current. The merge step picked it because the higher weekly number always won.
The table is built by a script, and the script reported what its source said. The source was frozen. The fix, recorded on September 26, 2026, had three parts: ignore any cache older than 30 minutes, let the fresher reading win, and show the age of every reading, tagged stale after one hour.
I take two rules from that. A broad set of sources raises the odds that one of them is wrong without anyone noticing. And every number an assistant gives me should arrive with a timestamp, so I can see staleness myself.
How do two assistants hand work to each other?
Sometimes a question crosses the line between the two domains. One assistant can then pull the other into the conversation by mentioning it, and the second answers from its own sources.
That handoff has failed. In August, Tao wrote a plain-text mention of another assistant in a shared channel and that assistant never answered. Two separate breaks sat behind it: the directory of assistants had never been loaded into Tao’s context, and Tao’s own messages were being dropped by a loop guard because Tao was missing from an allow list. The repair on August 11, 2026 added an audit script that checks those preconditions, and a test mention written by Tao got an answer back in 4 seconds. The same log entry listed further gaps still open that day, so a handoff that gets no reply is a defect to chase.
What am I measuring, and what have I measured?
Nothing yet. These are success criteria I am setting, not results I have measured:
- little stale or missing source data
- few repeated briefings
- actions that actually close
Scoped source connections and memory exist for both assistants. A baseline for completed and blocked handoffs, repeated briefings, source freshness and corrections of stale context does not. Without a baseline I cannot say whether the setup saves time, and I would rather leave that open than invent a figure. More agents alone only make a bigger diagram.
Where does this setup fail?
Context goes stale, as the usage table showed. The state file Tao reads can be out of date too, which is why the charter says to check it against the source systems before advising. The charters also name the cases where an assistant stops. For Thy that means anything that commits my time or money, sends in my name, books travel or speaks for me externally: draft it, get approval, do not send. Tao confirms before anything medical, financial, legal, irreversible or sent externally. The same line between acting and asking runs through my risk tiers for agent actions.
Who does this fit, and how do I start in 15 minutes?
It fits a leader whose week splits into clearly different domains, such as health and training on one side and logistics and relationships on the other, and who already has data worth connecting. If your whole job sits in one domain, one assistant with fewer sources may be enough.
Pick two domains from your own week. For each, write a charter of about ten lines: the objective, the three or four sources it may read, what it may do without asking and what always needs your approval. Add a single rule that every factual answer states its source and how old it is. After a week, count the times you had to brief an assistant twice, and note any answer built on data that had gone stale. Those two counts are your baseline.
You can read the series hub at Compounding CEO, and the practice behind this page is Give each AI assistant a domain and fresh sources.
Frequently Asked Questions
Why use two AI assistants instead of one?
Each job needs different sources and a different memory. My head coach works from training and recovery data, and my executive assistant works from commitments, travel and follow-ups. Keeping them apart means each assistant draws on sources that fit its job and neither carries the other's clutter.
Can an AI assistant have full context?
No assistant has it. Sources go stale, connections break and some information was never captured. I avoid the phrase because pretending otherwise is how stale information ends up driving a decision. I rate fresh sources above broad ones.
How do I know an assistant is working from stale data?
Make every claim carry its source and its age. In one case a usage table showed an account at 93 percent for 12 hours while the real figure was between 57 and 73 percent, because a cache outlived the login that wrote it. The fix was to ignore old caches and to show the age of every reading.
What should an AI assistant never do on its own?
Anything that commits your time or money, sends a message in your name, books travel or speaks for you externally. My executive assistant drafts those and waits for approval. My head coach confirms before anything medical, financial, legal, irreversible or sent externally.
Ready to Transform Your AI Strategy?
Get personalized guidance from someone who's led AI initiatives at Adidas, Sweetgreen, and 50+ Fortune 500 projects.
