One manager, named
Every employee reports to a specific person. That person sets its rules, approves its work, and is who an escalation goes to. No orphaned agents.
Offering 2 · Advanced build
Unburden your team
AI Employees work on top of your foundations layer. We call this the Nexus.
Nexus noun “nek-səs”
a central link, connection, or focal point where multiple things, ideas, or parts of a system come together
Nexus is where we bring together your employees and the power of AI to enable new, profitable scale.
Nexus is the deeper build. It assumes your team already has AI connected to its tools and is running shared skills and rules — if you're not there yet, we'll say so and point you at Foundation first.
When is it time to hire an AI employee?
Outgrowing foundations.
Teams that are early with AI should not start here. But once your whole team has AI wired into everything and is genuinely using it, the next limit shows up fast: one generic assistant holding everyone's context, everyone's tools, and everyone's rules at the same time is worse at every individual job than a focused one would be.
Nexus splits it up the way a company does. Each AI employee has a role, and only the skills, connections, memory, and rules that role needs. Better focus, cheaper to run, and far easier for a manager to reason about.
Everything below is the actual specification: how an employee is defined, where it runs, how work moves, and what stops it.
The org chart
Roles that report to a person, not chat windows that report to nobody.
Every AI employee has a name, a job, a set of things it owns, and one human it answers to. If you can draw it on your org chart, you can manage it like the rest of your org chart.
Head of Marketing
Human manager
Sets the rules, scopes the work, and is the only one who closes a task
Ellery
Content strategist
Owns the editorial calendar, briefs, long-form drafts
Ridge
Campaign analyst
Owns attribution, spend reporting, data validation
Sable
Social manager
Owns the publishing queue, community replies, asset library
Nova
Reviewer
Owns review. Nothing reaches the manager without passing here
Dotted line — Nova reviews the other three. Nobody manages anybody: work that isn't yours goes back up to the human, never sideways.
Each role carries only what it needs — which is why it's better at the job, cheaper to run, and easier to audit than one assistant carrying everything. Four is an example, not a limit; departments nest the same way.
01
02
No sideways authority
An AI employee can't assign work to another one. A task in the wrong lane comes back to the human with an explanation, which is a good outcome, not a failure.
03
One reviewer, always
At least one employee's job is checking the others. It's the cheapest headcount on the chart and it's the reason your attention gets spent last.
The employee definition
An AI employee is a file, and the file is the source of truth.
Not settings buried in a vendor dashboard. Each employee is a definition we write with you and keep in version control — so you can read it, diff it, review a change to it, and roll it back.
RG
Ridge
Campaign Analyst
$400/wk
Spend cap
Reports to
Head of Marketing · reviewed by Nova
Owns
AttributionSpend reportingData validation
Accounts
EmailSlackCRM seat
Connections
Warehouse · readCRM · readAds · readDrive · read/write
Skills
Validate report dataWeekly spend rollup
Model
Auto · routes to the cheapest model that can do the job
What it can never do
- Never writes to the warehouse. Read-only, always.
- Never reports a number it did not recompute from source.
- If a figure cannot be reconciled, it stops and asks instead of estimating.
Redacted example. The rules are written in plain language on purpose: it is the document a manager actually reads, edits and signs off on, not a config file.
What each part does at runtime
| Part of the definition | What it controls | Why it is separate |
|---|---|---|
| Role & owns | The job description, and the list of things this employee is accountable for | Work outside the list gets handed back instead of guessed at |
| Reports to | Which human approves its work and receives its escalations | An agent with no named manager is an agent nobody reviews |
| Accounts | Its own email, chat handle and seats — provisioned and revoked by your IT | Every action is attributable to it, not to whoever’s login it borrowed |
| Connections | The systems it can reach, at the access level it needs | Scoped per role, so the finance employee’s reach is not the social employee’s |
| Skills | The specific workflows for this job, versioned and shared where it makes sense | A focused catalog beats one library of everything the company does |
| Rules | What this role must never do, and what to do instead | Prose a manager can edit, reviewed in version control like code |
| Memory | Its own notes — what it learned, what it was told, what it already tried | Nothing leaks between roles and nothing bloats into one shared brain |
| Reviewer | Which other employee checks its output before a human sees it | The review gate is configuration, not a habit someone has to remember |
| Model & budget | Which model tier it routes to, and the ceilings on one run and on a week | Cost is a property of the role, so a cheap job cannot quietly spend like an expensive one |
Changing an employee is a reviewed diff: a change, a second pair of eyes, a deploy, and a revert if it was wrong. That is the whole reason the definition is a file.
Architecture
Five layers, and you own the ones that matter.
Nexus runs on infrastructure you control, against the systems you already run. It is not a seat in somebody else’s product.
Layer 5
People
You and your team
Scope work, set rules, approve what comes back. The only layer that closes a task.
↓
Layer 4
Surfaces
Slack, email, your project tool, your own app
Where work is assigned and answered. No new inbox — the board, the chat and the mail thread are all the same employee with the same memory.
↓
Layer 3
Orchestration
The dispatcher — the one writer of task state
Assigns, claims, serializes and times out every run. One authority moves a task, so two parts of the system can never disagree about what state it is in.
↓
Layer 2
Employee runtime
One isolated session per run
Loads that employee’s rules, skills, memory and credentials — and nothing else. Narration, tool calls and cost stream to the audit log as it works.
↓
Layer 1
Models & systems
Model-agnostic harness, your connected systems
Any frontier model, more than one vendor, or self-hosted. Underneath: your CRM, warehouse, drive and internal tools, reached with per-role credentials.
Credential isolation
A session can only use the accounts its own employee holds. Not “can only see” — can only use. It fails closed: an identity that cannot be resolved gets nothing, never everything.
Harness-agnostic
The engine underneath is a swappable adapter. Adding a vendor is one adapter, not a migration — which is how you avoid betting your org chart on one company’s pricing page.
Runs where you want
Your cloud, your server, or ours. Task history and memory live in a database you can back up, read and export without asking us.
The board
Work is a card with a state, an owner and a reviewer.
Managing a fleet is a management problem, not a chat problem. So the interface is the one you already use to run people: a board, with lanes, assignees, and a queue you can re-order by dragging.
Queue · 5
Re-run validation on 40 archived records
RG Ridge · review: Nova
WeeklyTop of queue
Compile the cross-partner newsletter
EL Ellery · review: Nova
Attempt 2 of 3
In progress · 2
Reprocess 200 flagged records
RG Ridge · 4m 12s
Running$1.84
Rewrite plain-language summaries for the top 10
EL Ellery · 1m 03s
Running
AI review · 2
Quarterly performance report
NV Nova reviewing · by Ellery
Gate 1
Update processing thresholds
NV Nova reviewing · by Sable
Gate 1
Manager review · 2
Monthly revenue-impact report
RG Ridge · passed Nova
Gate 2 — you$3.10
Raise paid-search budget on the partner campaign
SB Sable · needs a decision
BlockedWaiting on you
Done · 41
Q3 channel mix brief
EL Ellery · approved Tue
Reconcile August ad spend
RG Ridge · approved Mon
Representative UI. Every card carries its employee, its reviewer, the cost so far, the attempt count, and whether it is waiting on a person. Recurring work re-creates itself on a schedule.
Opening a card
A task is not a black box that returns a result. It is a thread: every state change, every tool call, every dollar, and the one line the employee signs off with.
Monthly revenue-impact report
Manager review
Schedule
Card created — recurring, first Monday
Dispatcher
Assigned to Ridge · reviewer Nova
Dispatcher
Queue → In progress
Ridge
14 tool calls · 6m 31s · $3.10
Ridge
Run ended — returned done
Dispatcher
In progress → AI review
Nova
Recomputed the figures from the source audit logs
Dispatcher
AI review → Manager review
You
Reviewing
RESULT: pass — figures reconcile within 0.4%; one wording change
Every state change stamped with who made it. Not “the AI did something” — a name, an action, a time and a cost.
The hand-off is one line, on purpose
An employee ends a run by stating a verdict in a fixed form: done, or blocked with the question. A reviewer states pass, fail with what is wrong, or blocked.
That line is the whole contract. The employee never moves its own card; the dispatcher reads the verdict and moves it. An agent that can promote its own work has a polite suggestion of a review gate.
- No verdict is not a pass. A review that ends without one costs an attempt and the card stays where it is
- Blocked means stopped. The run ends, the lock releases, and the question goes to the manager — it does not guess to keep the lane moving
- Three failed attempts and the card blocks itself with the reason, instead of burning budget in a loop
- A clean hand-off resets the count, so normal rework never trips the circuit breaker
The full state table
| State | Who holds it | What moves it on | Where it can go |
|---|---|---|---|
| Draft | A person | Scoped and assigned by a human | Queue |
| Queue | Nobody yet | The dispatcher claims it for a free employee — oldest first, unless pulled to the top | In progress |
| In progress | One AI employee | The run ends with done or blocked | AI review, or Blocked |
| AI review | A second AI employee | The reviewer returns pass, fail or blocked | Manager review, top of Queue, or Blocked |
| Manager review | You | You approve, or send it back with notes | Done, or top of Queue |
| Blocked | You | Your answer on the thread lifts it | Top of Queue, with your reply attached to the next run |
| Done | Closed | — | Only a human puts a card here |
One run per employee at a time, and reviews are dispatched before new work. Drag a card somewhere the rules do not allow and it normalizes into the nearest legal state. A board should behave like a board.
Two review gates
Nothing reaches you until another employee has checked it.
The stage most AI tools are missing sits between “done” and “your problem.” Work goes to a second AI employee for review first, and only reaches a person once it has survived that.
Draft
Scoped by a human
Queue
Assigned & ordered
In progress
AI employee working
Gate 1 — AI review
A second AI employee
Gate 2 — you
Last, and only once
Your attention is the scarcest thing in the system, so it is spent last — on work that has already been checked by someone whose entire job was checking it.
What gate 1 can return
PASS
It holds up
The reviewer re-derived the claims and agrees. The card moves to your lane with the reviewer’s note attached, so what you read is a second opinion, not a first draft.
FAIL
It goes back, with notes
Straight to the top of the queue with what is wrong. The next run gets those notes in its brief, so it picks up where the last attempt stopped instead of starting over. You are not involved.
BLOCKED
It needs a decision
The reviewer hit a question only a person can answer. The run stops, the question goes to you, and your reply re-queues the card with the answer attached.
Rework between two AI employees is the cheapest loop in the system and the one you never see. That is the point: the expensive reviewer is you, so gate 1 exists to waste as little of gate 2 as possible.
Why a second employee, not a second prompt
They check each other’s work.
A reviewing employee does not just re-read the draft. It has its own tools, credentials and context, so it goes back to the source data, runs the numbers itself, and disagrees when they say something different.
This is the practical argument for a fleet: asking a model to double-check itself gets you the same blind spot twice. A different employee with different access catches what the first one never would.
- Different credentials — the reviewer reads the source system directly, not the draft’s summary of it
- Different rules — a reviewer’s rules are about verification, not production
- No editing — a reviewer returns a verdict. It never quietly fixes the work and passes it
Quarterly performance report
Gate 1 — transcript
Ellery Content strategist · AI employee
Quarterly performance report drafted — three findings, three recommendations, each one sourced to the campaign tables. Run ended; handing off to gate 1.
Nova Reviewer · AI employee
$ python3 scripts/validate_report_data.py --q3
Headline totals reconcile against the source audit logs within 0.4%, so findings #1 and #3 stand. Finding #2 does not. The dip Ellery reads as a quality regression is concentrated in the two weeks immediately after the mid-quarter config rollout — that is a deployment artefact, not a regression, and the report recommends the wrong fix on the back of it. Failing it, with the rollout timestamps attached.
Dispatcher The system · not an employee
Gate 1 returned fail, so the card goes to the top of the Queue with Nova’s notes attached — Ellery’s next run opens with them. Attempt 2 of 3. Neither employee moved the card.
RESULT: fail — finding #2 is a rollout artefact, not a regression
An AI review that changed the conclusion — before a person spent a minute on it.
Dedicated, not shared
They are real employees. They have their own accounts.
A Nexus employee does not borrow a human’s login. It gets its own email address, its own chat handle and its own seat in your CRM, provisioned, permissioned, and revocable like anyone else on payroll.
Ridge
Campaign analyst · AI employee
ridge@yourcompany.com
@ridge
Salesforce seat
HubSpot user
Drive account
Board member
Its own identity in every system — so every action it takes is attributable to it, not to whoever’s credentials it borrowed.
This is the part most AI tools skip, and it is the part that makes the difference between a clever assistant and something you can actually put on the org chart.
- Own accounts — real logins in Gmail, Slack, Salesforce and everywhere else it works, provisioned by your IT the same way a new hire’s are
- Siloed memory — it remembers its own work and its own context. Nothing leaks between roles, and nothing bloats into one shared brain
- Its own permissions — scoped to the job. Your finance employee can reach the billing system; your social employee cannot
- Its own rules — what this specific role may never do, set by the manager it reports to
- Its own skills — the workflows for its job, not a catalog of everything the company does
- Its own training — onboarded on your systems, your voice and your process, and retrained as they change
Because every action runs under its own identity, your existing audit logs, access reviews and offboarding process already work. Turning an AI employee off is the same as turning any employee off.
Where they work
No new inbox to check.
Nexus employees work inside the tools your team is already in. Assign work on the board, ask a question in Slack, forward an email — it is the same employee, with the same memory of what it is working on.
Slack
Ellery #marketing
Q3 brief is ready for review — Nova validated the numbers, one recommendation changed.
Project tool
Reprocess 200 flagged records assigned
Assigned to Sable and moved to In progress, with a time estimate in the comments.
Email
Ridge ridge@yourcompany.com
Replying on the forwarded partner thread, with the three action items pulled out.
Web & mobile
Nexus your own app
Every task, employee and audit trail — on a laptop or a phone.
Work arrives three ways
A person files it
Dragged onto the board, asked for in chat, or forwarded as an email. Ad-hoc work, scoped by whoever wants it.
A schedule files it
Recurring cards that re-create themselves, the weekly rollup and the monthly report, on your timezone, with a run history you can audit.
A system files it
A webhook from your own tools: a deal closes, a form comes in, a threshold trips, and the right employee picks up a card for it.
Two things you only get from a fleet you own.
Mini apps
Need a tool? They will build it.
Some work needs more than a conversation. Nexus employees can build and host small internal tools for your team — no engineering ticket, no new vendor, permissions handled on the same platform.
Expense approvals
Pipeline dashboard
Lead scoring
Contract sign-off queue
Running lean
Your subscriptions do the work, not metered API pricing.
Metered API access is priced per token, and it adds up fast at real usage. Nexus runs on the flat-rate subscriptions your team already pays for instead, and only reaches for metered API when a subscription genuinely cannot do the job.
Metered APIExisting subscriptions
Auto picks the model. Every run routes through Auto, which sends the job to the cheapest model that can actually do it, and steps up to a frontier model only when the work calls for it.
Context stays lean. A long-running thread gets compressed and handed off with only what the next step needs, instead of dragging the whole history along and paying to re-read it every time.
Subagents take the side work. Research, verification and narrow sub-tasks run in their own isolated calls, so the main job never pays premium-model rates for work a cheaper pass could do.
Every run reports what it cost, per card and per employee.
Operational spec
The boring questions, answered up front.
What it does when it fails, what it costs you in attention, and who can see what. These are the things that decide whether a fleet survives contact with a real company.
| Concern | How Nexus handles it |
|---|---|
| Concurrency | One run per employee at a time. Reviews are dispatched before new work, so nothing finished waits behind something starting. |
| Failure | A crashed or capped run counts as an attempt. Three in a row and the card blocks itself with the reason, rather than looping. |
| Runaway cost | Per-run turn and dollar ceilings, plus rolling limits per employee. A run that hits a ceiling stops and reports instead of continuing quietly. |
| Restart | Held cards are re-queued without charging an attempt, so a deploy or a reboot never eats someone’s work. |
| Audit | Every state change, tool call, narration line and cost is stored per card, attributed to a named employee, and exportable. |
| Data location | Your infrastructure or ours, your choice. Task history and memory sit in a database you can back up and read directly. |
| Access control | Per-role credentials, enforced at the moment of use and failing closed. Employees cannot reach each other’s accounts or memory. |
| Offboarding | Revoke the accounts. It is the same process as a human leaving, because it is the same kind of account. |
| Changing an employee | Edit the definition or the rules file. It is a reviewed diff in version control — deployable, and revertable. |
| Human time required | Scoping work, answering blocks, and gate 2 approvals. Rework between employees never reaches you. |
How Nexus is different
Against what you are running, and against the closest thing shipping.
Most teams are in the first column. Grok Bot is the nearest product to Nexus shipping today: persistent AI agents that sign in to your tools and work on their own. It is real and it works.
| One shared assistant What most teams run today |
Grok Bot xAI |
Nexus — AI Employees Corvino Labs |
|
|---|---|---|---|
| What it is | One assistant the whole team shares | Persistent bots, each on its own cloud machine, coordinating in a group chat | Named roles on your org chart, each reporting to a human manager |
| Identity | Runs as whoever is typing | Signs in to your existing apps and operates them | Its own accounts and seats — own email, own chat handle, own CRM login, provisioned and revoked like any employee |
| Org placement | None. It is a tool, not a role | A bot you configure, outside the reporting structure | On the chart, with a named manager, a job description and a list of what it owns |
| Memory | One context everybody shares and nobody curates | Per-bot context | Siloed per role — nothing leaks between employees, nothing bloats into one shared brain |
| Skills | Each person builds their own, from scratch | Per-bot instructions | Versioned library, grouped by role, so each employee stays focused and easy to hand off |
| Rules & guardrails | Whatever each person remembers to type | Per-bot instructions; enterprise access and audit controls | A rules file per role, set by that role’s manager and reviewed as a diff in version control |
| Tool access | Whatever ships in the box | Signs in to your apps and operates them like a user | Off-the-shelf connectors, plus custom ones we build for your internal and no-API tools — scoped per role |
| Quality control | You, on everything, every time | Runs until it needs your sign-off | Two gates — a second AI employee re-derives the work against source data before a human ever sees it |
| When it is unsure | It guesses, fluently | Pauses for your sign-off | Stops, states the question, and waits. The card blocks rather than proceeding on an assumption |
| When it fails | You notice, eventually | Retries in its own session | Attempt counted, notes attached, back to the top of the queue — three in a row and it escalates to a person |
| Managing the work | A list of recent chats | Its own chat surface, with sign-off prompts | A board with states, owners, reviewers, attempt counts and per-card cost |
| Where the work lands | The vendor’s chat window | Its own chat surface | Your project tool, Slack and email — plus an app of your own where it helps |
| Recurring & triggered work | The vendor’s scheduler, with its limits | Runs continuously on its own machine | Scheduled cards on your timezone, plus webhooks from your own systems, both with a run history |
| Model | One model, one vendor | Grok only | Model- and harness-agnostic, including self-hosted |
| Cost shape | Subscription, plus overages, plus cloud costs | Roughly $120 per bot, per month | Uses the flat-rate subscriptions you already pay for, with automatic routing to the cheapest model that can do the job |
| Cost visibility | One invoice, no attribution | Per bot, per month | Per run, per card and per employee, with ceilings you set |
| Audit & offboarding | Every action looks like the human’s | Enterprise access and audit controls | Every action under its own identity, so your existing audit and offboarding already cover it |
| Internal tools | Throwaway apps that run on one laptop | Can write code for you | Builds and hosts small internal apps, with permissions, on the same platform |
| Where it runs | The vendor’s cloud | One cloud machine per bot | Your infrastructure or ours, with the data in a database you control |
| Who it is built for | Individuals | General knowledge work, any company | Your company — roles and rules shaped to how your team actually works |
Comparison based on publicly available information as of September 2026. Grok Bot is a trademark of xAI. Both are shipping products; the differences above are design choices, not roadmap promises.
Is your team ready
Nexus is the advanced build. Not everyone should start here.
It is live and we will build it for you today. But it pays off for teams that have already done the groundwork — and we would rather tell you that now than six months in.
Nexus is a fit if…
- Your team already uses AI daily and is hitting its ceiling
- Your tools are connected and people trust the output
- You have distinct roles that need distinct skills and rules
- Somebody internally is accountable for how AI gets used
Start with Foundation if…
- AI still cannot reach most of your systems
- Only one or two people are getting real value from it
- Nobody has written down what it is allowed to do
- You are not sure yet which work is worth handing over
Let’s find out which one you need.
Thirty minutes on how your team works today. We will tell you honestly whether you are ready for Nexus or better served starting with Foundation.