Architecture
This is the how. A plain walk through the way TagTeam works, end to end. The goal of every piece below is the same: give you your time back, so you can work effectively without staring at a screen.
The linchpin: your Clone
You talk to Claude. Your Clone does your work.
Your Clone is the heart of the whole system. A Clone is Claude equipped with a strong setup, world-class Skills, and Memory, running on your behalf. There is one Clone per person. It knows you, acts as you, and dispatches Agents to get things done while you are off the screen.
Everything else here, the Agents, the Skills, the Memory, the daily cycle, exists to build, equip, and keep sharpening your Clone.
The #1 metric: Human Hours Saved
We measure success with one number above all others: Human Hours Saved. It is the realistic best estimate of the time the system gave back to you by doing work you would otherwise have done yourself.
It counts honestly. That means it includes the time you would otherwise spend working with the AI: the prompting, the supervising, and the waiting, not just the task itself. It is reported first, ranked above tokens used and every other cost gauge. Every cycle should add to the number.
The bar for good: the 4 E's of Excellence
Every Agent, every Skill, and every output is judged against four words. Miss any one and it is not done.
- Effective: it achieves the result you wanted, not just "ran without error."
- Efficient: it gets there with the least wasted effort and compute, never by cutting context that would have made the result better.
- Ethical: it puts security first and does the right thing, always.
- Easy: it is calm to drive and clear to read, with nothing for you to babysit.
The team: your Clone plus five Agents
An Agent is an autonomous worker that owns one domain completely. The system runs on a few powerful Agents with world-class Skills, not a sprawl of narrow bots.
The roster is your Clone plus five Agents, and it is capped at six. Your Clone is the leader: it knows you, owns the result, and dispatches the others to their lanes. Three are work agents your Clone sends to do the work. Two are framework agents that build and run TagTeam itself.
| Member | What it owns |
|---|---|
| Your Clone (one per person) | The leader. Knows you, acts as you, owns the result, and dispatches the work agents and approves their output. |
| Maker Agent | Makes the system's parts: new Skills and Agents, or rebuilds existing ones from the ground up. It is the only Agent that creates Agents. |
| Manager Agent | Manages the recurring real-world work: inbox, calendar, projects, sending, booking, and posting. The system's hands, held to the tightest checks. |
| Monitor Agent | Watches how things perform over time, and does the research, the analysis, and the synthesis, so your Clone gets the conclusion, not the raw load. |
| TagTeam Customizer | Onboards you, equips your Clone, and keeps tuning the system with you over time: your voice, your gates, your sources, your goals, what to never touch. |
| TagTeam Updater | Keeps the live system current on its own: tends Memory, tightens what exists, runs the daily cycle, and produces your Daily Digest. |
You never name the work agents yourself; your Clone dispatches them. A new job does not become a new Agent. The system grows by deepening an Agent with a new Skill, which is what keeps the roster small and the team powerful.
Behind it all, Claude is the chat surface, the window where you talk. It reads your Clone's Memory, but only to know you.
Skills
A Skill is a packaged, reusable capability that an Agent uses to get something done. Capability lives in Skills; orchestration lives in Agents. The rule is few powerful Agents and world-class Skills. When a new capability is needed, the answer is to add a Skill, not to spin up another narrow Agent. Skills are how the system grows; Agents are how it operates.
How it reads your world: source synthesizers
Your Clone gets sharper by reading the world you already work in. It does this through a family of source synthesizers, one for each kind of source you connect:
- Calendar
- Documents
- Meetings (read from the actual transcripts, with the most recent meetings read in full)
- Your AI chats
- Social
- Web
- Your project tools
Every synthesizer runs the same three-stage gate, every time:
- Gather: read the sources, after sensitive content has already been filtered out, and assemble candidate learnings. Nothing is written yet.
- Draft: bundle the proposed changes into one clear package, each with its evidence and a confidence rating.
- Confirm: show you a short summary with the proposed changes, then stop. Nothing commits until you say yes.
The default output of any run is a staged draft, waiting for one decision. A single low-confidence hunch never gets memorialized, and your Profile is never silently overwritten.
Connectors: source and tool in one
A Connector is one live link to an outside service, like your email, calendar, or project tools. It is both how the system reads that source and how it acts on it: the same link that lets your Clone learn how you write email is the link it later uses to send email as you. One link, with the credential held safely outside chat and Memory.
The Memory model
Memory is the system's deepest piece. It is split into three stores so it is never one giant pile: one for you (your Clone), one for your company, and one for the system itself. Inside each store, the same three levels:
| Level | What it holds |
|---|---|
| Active | Your current operating picture: who you are, your goals, your plan, work in progress. Read first, every cycle. Kept short and sharp. |
| Adaptive | The capture layer: signals, open issues, and a rolling stream of what changed or needs a decision. Ephemeral by default; items fade if they are not confirmed. |
| Archival | Durable, confirmed truth. Rarely edited. Nothing lands here without your confirmation. |
Two files anchor it all: a holistic profile of you, and a parallel profile of your company. Every Agent reads the right profile first before acting on your behalf. The system keeps both current with regular alignment checks that ask: does the profile still match what the sources show? When something has drifted, the change is proposed for your confirmation, never written silently.
The Review Gate and the three approval tiers
Your attention is reserved for what truly warrants it. So every action runs through the Review Gate, the checkpoint that decides what just happens and what asks you first. The law has three tiers.
Above all three tiers sit the security absolutes. These are never approvable and never negotiable. Even a Tier 3 yes from you cannot override them.
- Filtered before it is ever read. Sensitive categories, financial, HR, legal-privileged, payment info, health, and personal life, are filtered out before content ever reaches the system, not cleaned up after. An excluded item simply never enters the pipeline.
- A final safety scan you cannot override. Before anything is saved, a secret-and-personal-data scan runs as a hard gate. If it finds something, the save is blocked, even when you already approved it. Approval clears taste and judgment; the scan clears safety, and safety wins.
- Never a password or a one-time code. The system never asks for, accepts, stores, or sends a password or a one-time passcode, ever. If a task seems to need one, it stops and has you enter it yourself.
- Off-limits means off-limits. Anything whose name contains the word "private," and any zone you mark as no-go, is never opened, read, or accessed, under any circumstances.
With those absolutes in place, ordinary work sorts into three tiers:
- Tier 1: just happens. Routine, reversible background work: notes, captures, housekeeping, reading, preparing drafts. No report needed, and every write is reversible.
- Tier 2: happens, then is reported. Real work the system did that you would want to know about: completed tasks, notable findings, drafts ready for you. Done on its own, reported in your Daily Digest.
- Tier 3: asks first. Anything foundational, irreversible, or outward-facing: changes to your profile, goals, or confirmed truths; sending, posting, booking, or deleting anything outside; creating anything new; or anything touching money or other people's time.
The test is magnitude. New or fundamental means ask first; small, reversible, intent-preserving means it can run on its own. When in doubt, the system goes up one tier.
Security, from day one
Security is foundational, not a feature added later. Data flows only where it must, and only as much as it must. The filtering, the final scan, and the two absolutes above all run from day one, on every read and every save. Secrets like passwords, keys, and tokens never live in chat, Memory, or anywhere the Agents can see; they stay locked away on the host. And when access is missing, the system says so plainly rather than working around it.
The four things the system does: Run, Record, Report, Request
When something surfaces, a request, a signal, an idea, a problem, the system handles it with four moves, in any combination:
- Run: do the work now, schedule it for later, or queue it for the next cycle.
- Record: write what happened to the right place in Memory.
- Report: tell you what happened, when no answer is needed.
- Request: ask you for what is needed to proceed. A Request is never empty-handed: it always brings something concrete, a question to align on, an approval to grant, or finished work with refinements proposed.
The daily cycle, the Pulse, and your Daily Digest
The system's heartbeat is the daily cycle, owned by the TagTeam Updater. Each run reads the current state, checks everything for freshness, runs the alignment check that keeps your profile matched to reality, consolidates confirmed observations toward durable truth through the gate, and surfaces what changed.
This regular cadence is called the Pulse: capture, route what changed, check for contradictions and anything blocked on you, then produce the digest. Behind it sits the Adaptive engine, the always-watching layer that captures signals as they land. The engine is what watches; the Pulse is when it runs. You can run the Pulse on demand any time by saying update.
The result is your Daily Digest: a single, calm, once-a-day summary. It leads with what is waiting on you, gives you a clean checklist of what got handled, and brings any real decision with options and a recommended path, never a bare question. You stay in the loop on what you would weigh in on, and you are never buried in housekeeping.
When you want something done, you use Tag-In: you tell Claude what you want, you tag in your Clone, and you close the window. Your Clone goes to work and reports back when it is done. Those are the only three words you ever need: Tag to set your Clone to work, Customize to teach and tune it, and Update to run the Pulse right now.
What's live today
We design for always-on, and we tell you honestly what runs today. The system is built for continuous, always-watching operation, and the engine for it is already in place.
Today the Pulse runs live, every hour, server-side as a Deployment on Anthropic's platform: no temporary scheduling bridge left in the loop. Your Daily Digest composes at the hour you choose. Connectors attach through your own vault, and live capture through them is being proven now; until it is proven end to end, we will not describe the loop as always-watching, because it is not yet. The moment it lands, the cadence tightens with no change to how any of this works.