Type a sentence. It becomes tool calls.

What is the Nerxus One agent screen?

The Nerxus One agent screen is the chat panel where you type a sentence and Nova turns it into specialist tool calls, then comes back last with one reply. It is the same orchestrator that automation can wake on a trigger. The entry point changes. The loop does not: classify, assemble scoped memory, evaluate guardrails, dispatch, synthesize.

The buyer is the marketer. You type. The dentist does not. The panel lives in the authenticated app shell, next to the module you are already in. Ask from Leads, Radar, Outreach, Clients, or Mission Control. The specialist still writes to its own module. The screen does not become a second copy of the database.

Hero copy on this site shows a typical first sentence: find dentists in Austin with no online booking and a dropping rating. The visible trail is Scout, then Radar, then Forge if you asked for a DEMO. That trail is the product. A typing animation that never names a tool is a demo of tone.

The panel is available on every module page for a reason. You should not have to leave Clients to ask Steward for a recap, or leave Leads to ask Scout to score an imported list. Context can include the business you already have open. If the subject id is present, memory assembly prefers that row. If it is missing, Nova should ask which business before Relay drafts.

How does a typed sentence become tool calls?

A typed sentence becomes tool calls because Nova tokenizes the text, scores it against every registered tool's keywords, picks a winner or asks a question, then runs that tool only after the guardrail verdict. The runtime names the steps VALIDATE, CLASSIFY, ASSEMBLE CONTEXT, DISPATCH, SYNTHESIZE. Auth and tenant scope happen upstream. Classification is the first thing Nova owns.

Each tool in the registry carries a name, an owning agent, a keyword list, a description, and a safety tier. scout.search_leads is not a vibe. relay.send_email is not a vibe. The sentence has to match a tool or the screen stops and asks. That is cheaper and safer than letting a model invent an action the registry never declared.

After a tool runs, Nova writes an episodic memory row and emits agent.action_taken. The activity feed and the audit log see the same summary. If the verdict needed approval, the reply says so and the item lands in Approvals. You do not have to guess whether the sentence already mailed someone.

A tool that fails still comes back through Nova. Missing Google Places on a search should name Settings, Connections. An unverified domain on a send should name senders. A locked SMS intent should name Forms and the inbound line. Silent failure is how a chat UI trains you to retry until something unsafe works.

How does Nova classify intent on the agent screen?

Nova classifies intent in two tiers: deterministic keyword overlap first, then an LLM fallback only when the score is ambiguous. Tier 1 lowercases the sentence, strips punctuation, and scores the fraction of a tool's keyword phrases that appear. A high score routes directly. That path costs no tokens and finishes in milliseconds.

A mid score can still route and flag. A low score does not pick a send to be helpful. If an LLM classifier is wired, Nova hands it the tool names and descriptions and accepts a tool only if that tool exists in the registry. A model that names a tool you do not have is ignored.

Below the floor, Nova asks one clarifying question. The stock question lists the jobs it can route: finding leads, diagnosing a business, outreach, reports, builds, reputation, or clients. It does not pretend it can cold-call. Compliance intents, such as text the whole list, are refused before any of those routes run.

What happens when Nova is not sure which tool to run?

When Nova is not sure which tool to run, it asks one question or shows you the playbook it would create, instead of dispatching a send. Guessing is how a marketer wakes up to a draft that already left. The classifier would rather look slow than look autonomous on a weak match.

You can also be specific. Name the agent. Name the metro. Name the detector. Find dentists in Austin with no online booking is a Scout sentence. Draft outreach for the businesses Radar flagged this morning is a Relay sentence. Who is about to churn is a Steward sentence. The more the sentence already sounds like a tool, the less the fallback model has to do.

Slash-delegate is the other escape hatch. If you already know you want Forge, you do not need Nova to rediscover that from a paragraph. The roster page exists so you can learn the names. The screen exists so you can use them without walking every sidebar child.

Why does Nova come back last on every request?

Nova comes back last because it is first-in and last-out by rule: it opens the request, it never mutates module state itself, and it synthesizes the specialist results into one reply. Specialists do the domain work. Nova talks to you.

That split is what keeps a multi-step ask honest. If you asked for a search, a diagnostic, and a draft, you should see three tool trails and one closing paragraph, not a mash that hides who wrote the email. The closing paragraph can link into Leads, Radar, Approvals, or a playbook card. It should not invent a finished overnight run.

Coming back last is also how memory stays scoped. Nova assembles the minimum memory for the specialist it is about to call, tenant to subject to memory. Scout does not need Steward's invoice thread. Relay does not need a raw Places dump. Last-out is when the episodic row is written, so the next sentence can see what already happened.

What can I type on the Nerxus One agent screen?

You can type any job the registry already owns: hunt a vertical and metro, diagnose a saved business, park a deal, draft signal-led mail, compose a report, open a DEMO, draft a content week, check reputation, pull a client, pause a playbook, or ask what is waiting on you. If the job is not a tool, Nova will say it cannot find a capability yet.

Automation sentences have working cards in the panel. What did you do last night returns a briefing card with three sections. Turn off the storm playbook returns the Storm-Event Trigger card and asks you to confirm the pause. Which prospects are growing returns the health board inline, bands only. Anything waiting on me returns the approval queue with Approve, Edit, and Skip. Run the overnight loop as a dry run stages a simulation that touches nothing real.

Watch my Austin dentists for storm damage composes a scoped Watchtower in the chat, with a matches-about-N preview, before you commit. Build an ICP for med spas in San Diego that skip online booking carries the sentence into the five-step ICP review. The blast radius stays on that screen. Activation stays yours.

Does typing on the agent screen send email by itself?

Typing on the agent screen does not send email by itself. A sentence that wants a send is classified toward Relay, then the gate runs, then Supervised queues the draft. Autonomous mode can still stop at the same gate if the domain is unverified, the window is closed, the cap is hit, or the playbook says first-run approval.

The reply text is supposed to be boring. Nova: this needs your approval before it sends. I've queued it in Approvals. That sentence is the product working. A cheerful I already reached out is a defect unless activity shows a Relay send with a pass verdict.

Client recaps and invoice mail follow the same rule after the sale. Steward drafts. Relay sends. The agent screen does not grow a quiet mailbox for retainers. If you want mail to leave, you approve it, or you later write a standing rule that names that exact send. Until then, the queue holds the artifact.

How do guardrails run before a tool call?

Guardrails run after classification and before the specialist tool, as a verdict: allow, queue, or block. The evaluator sees the tool, the current Agent Mode, spend caps, and the compliance locks. A blocked tool still writes an activity row so you can see what Nova refused.

Off mode stages and does not run. The panel attaches a note: Agents are off, I can stage this, nothing will run until you switch modes. Supervised lets read and draft work proceed and holds sends, publishes, and spend. The four locks refuse even if you type please or I own the list.

This is why the screen can feel strict on week one. The strictness is the same matrix you will later tune per agent and per action. The screen is not a separate, looser product. It is the same gate you will meet in Mission Control when a trigger tries the same tool at 2am.

What does a multi-step request look like on the agent screen?

A multi-step request looks like one sentence, a visible chain of specialist calls, and a last Nova paragraph that names what is saved, what is drafted, and what is waiting. Find, diagnose, and attach a DEMO is Scout, Radar, Forge. The business uuid is passed, not a provider id that dies when you rotate a key.

If a later step needs approval, earlier green steps can still finish. Scout can save. Radar can write findings. Forge can leave a DEMO. Relay waits. That is the sample-first motion without a premature send. You should be able to open the DEMO from the reply without wondering whether the owner already has mail.

If a later step is locked, the chain stops with the lock named. Text them after you find them does not become a search plus a blast. The search can run. The SMS cannot. Nova should say a discovered phone number is not consent, and then point at email, the inbound line, or a form checkbox.

Can I drive Mission Control from the agent screen?

You can drive Mission Control from the agent screen because automation is chat-driven, not only a settings page. The same Nova that routes Scout also pauses playbooks, renders the queue, and stages dry runs. The cards are real controls. Approve in the panel clears the item on the Approvals page.

What did you do last night is a briefing of the last recorded window. If nothing ran, the card should not invent scanned and sent counts from a seed tenant you are not. The governance plane is live. Overnight dispatch is the open half. A dry run is how you see the steps without pretending the night already happened.

Health questions return Healthy, Watch, or At-risk bands with drivers. They never print a client's revenue as a dollar. Pitch now, when it appears, starts the motion you already have: Pulse, Relay, Atlas. It does not print a prophecy about close rate.

What the Nerxus One agent screen will not do

The agent screen will not cold-call, will not cold-text a list, will not hide public review buttons, and will not send on an unverified domain. Those refusals are the same four locks as the rest of the product. Talking to Nova does not create a side door.

It will not invent a price, a customer, or a SOC 2 badge if you ask it to sound more enterprise. It will not treat Yelp or Tripadvisor as a live reply API. It will not claim Google or Facebook credentials have been exercised on your tenant if they have not. Unknown stays unknown.

It will not fulfill in five other tools for you. If you want the DEMO, Forge has to build it here. If you want the retainer to keep a history, Steward has to inherit the row here. The screen is fast because the modules exist. It is not a layer that makes missing modules appear.

How the agent screen differs from a generic chatbot

A generic chatbot answers in prose and hopes you will copy the answer into another app. The agent screen answers with tool calls that already know the tenant, the business uuid, and the gate. The prose is the wrapper. The row is the result.

You can still have a conversation. You can ask why a lead scored low, what a finding's assumptions were, or which renewal play Steward would start. Those answers should quote the record. If the record is empty, the honest reply is that the diagnostic has not run, not a guessed chair-hour number.

Marketers who want an always-on closer will find this screen conservative. That is the point of Supervised. The sentence is yours. The send is yours. The roster does the walking in between.

Who the agent screen is for

The agent screen is for a marketer who already knows the week and does not want to click six modules to start it. One sentence should open the hunt, the diagnostic, or the queue. Two to ten people can share Nova. Roles still decide who can approve a send.

If you hate chat UIs, you can ignore the panel and drive Leads, Radar, Outreach, and Clients as screens. The tools are the same. Nova is a door, not a requirement. If you only want a chatbot that writes blog posts, Muse exists, and publish is still gated.

If you want a zero-approval-click overnight machine on day one, this is the wrong page and the wrong product posture. Read guardrails and automation. Supervised is the default. Dry run is the first enable. Production overnight dispatch is not a claim we print.

How the agent screen handles Off, Supervised, and Autonomous

The agent screen handles Off, Supervised, and Autonomous by reading the same workspace setting Mission Control writes. Off attaches a note that work can be staged and will not run. Supervised lets green tools finish and holds amber ones with a queued reply. Autonomous still hits the four locks, the kill switch, caps, windows, and scope. The screen does not get a looser copy of the mode.

You should be able to see the mode without opening Guardrails. The top-bar pill is that signal. If you type a send while the pill is Off, Nova should not quietly flip the tenant to Supervised to be helpful. Mode changes are audited. A chat that changes mode without saying so is a defect.

Cards in the panel respect the same states. A dry-run card still runs in Off as a simulation. An Approve button on a queued send still re-checks the gate. A compose-Watchtower card still shows matches-about-N before you commit. The screen is allowed to be interactive. It is not allowed to be a second, softer policy.

How to start a session on the agent screen

You start a session by signing into the app after access opens, landing in the shell, and typing a sentence that names a vertical, a metro, and a problem you can fulfill. Find is the usual first verb. Prove and write come after a save.

If Google Places is missing, Scout cannot search. The screen should fail closed and point at Settings, Connections, not invent a list. If the sending domain is unverified, Relay can still draft. It cannot send. If Agent Mode is Off, you will get a staged card and a note.

Get started on this site is the waitlist until batches open. Login will point at the app when that URL is public. The marketing page you are reading is not the agent screen. It is the explanation of why the mock in the hero shows tool calls instead of a generic dashboard.

If you already have access, type the hunt you can fulfill this week. Name the vertical. Name the metro. Name the detector. Then stop. Open the saved rows. Run Radar on the three you would take. Ask for a DEMO only where Forge has a fulfillment action. Ask for a draft only after the sentence exists. That order is how the screen stays a router instead of a slot machine.

How does the agent screen assemble memory before a tool runs?

The agent screen assembles memory before a tool runs by taking the tenant, the subject business if one is open, and the minimum episodic and semantic facts that specialist needs. Scout does not receive Steward's invoice thread. Relay does not receive a raw Places dump. Atlas does not receive an internal note you marked off the portal.

If the subject id is missing and the sentence wants a draft, Nova should ask which business before Relay writes. A recap that guesses the shop from a similar name is how two books mix in prose even when RLS held. Context from the module you already have open is a hint, not a license to skip the uuid.

After the tool finishes, Nova writes an episodic row and comes back last. The next sentence can see what already happened. A nightly consolidation job is supposed to promote durable facts. Working memory is the current turn. None of that assembly gives the screen a side door around the gate.

Over-sharing is the usual defect. A reply that quotes a contractor aside from the hidden timeline is a leak in sentences. Scoped assembly exists so the specialist can work without sounding like your Slack.

What does a failed tool call look like in the reply?

A failed tool call looks like a named miss and a next setting, not a spinner that invents success. Missing Google Places on a search should name Settings, Connections. An unverified domain on a send should name senders. A locked SMS intent should name Forms and the inbound line.

The activity row still writes. Blocked is a decision. A weak classify that asked a question is also a decision. Silent failure is how a chat UI trains you to retry until something unsafe works. Retry should not get you a colder channel.

If a later step in a chain fails, earlier green steps can still stand. Scout can save. Radar can write findings. Forge can leave a DEMO. Relay waits. The last Nova paragraph should say what is saved, what is drafted, and what is waiting, without pretending the owner already has mail.

Seeded demo tenants can look healthier than a new book. If your Places key is missing, the screen should not borrow a sample list to look alive. Unknown stays unknown. A green check from nowhere is the same class of lie as a last-night briefing you did not run.

How do Approvals and the agent screen stay the same queue?

Approvals and the agent screen stay the same queue because a queued send is one item with two doors. Anything waiting on me renders the queue inline. Approve, Edit, and Skip persist. Resolving in chat clears the Approvals page. A toast that only hides the card is not a resolve.

Approve re-checks the gate. A domain that expired between draft and click still fails. Edit merges your reviewed input and then re-checks. Reject blocks the staged execution and down-weights that procedural pattern. Snooze stores a resurface date. The panel and the table should agree on age and reason.

Mission Control cards follow the same rule. Approve in the cockpit clears the chat card. The screen is not a second, softer policy. Off still stages. Supervised still holds amber. Autonomous still hits the four locks, the kill switch, caps, windows, and scope.

If two people share a tenant, decide who may click Approve. A queue anyone can clear is how a small team ships a collision. Roles still matter. Nova being shared is not the same as every seat owning the domain.

What context does the open module give Nova?

The open module gives Nova a subject hint and a default specialist neighborhood, not a second database. Ask from Leads and a hunt sentence should lean Scout. Ask from Clients and a recap sentence should lean Steward. Ask from Outreach and a draft sentence should lean Relay. The tool still writes to its own module.

If you have a business row open, memory assembly prefers that uuid. If you do not, Nova should not guess from a first name. Joe's downtown and Joe's south are two rows. A helpful mash is a leak.

The panel lives in the authenticated app shell on purpose. You should not leave Clients to ask for a recap, or leave Leads to score an import. You should also not treat the panel as a place to paste a patient list or a diner dump. Those files are not this product's ICP.

Slash-delegate beats a wrong neighborhood. If you are on Pulse and you want Forge, name Forge. Classification will still refuse a tool the registry does not have. The module you are looking at cannot invent an eleventh sender.

How should a new seat learn the agent screen without sending?

A new seat learns the agent screen without sending by staying in Supervised, typing Find and Diagnose first, and treating every send-shaped reply as a queue card. Name a vertical, a metro, and a problem they can fulfill. Save. Stop. Open the row. Ask for Radar. Ask for a DEMO only where a fulfillment action exists.

Off mode is the stricter classroom. Nova can talk and stage. Nothing runs. The note should say so. A trainer who flips the tenant to Autonomous so the junior can feel magic is borrowing the domain. Dry-run cards are the overnight lesson. They are not a night that mailed.

Teach the refusals in the same sitting. Text everyone on the list should fail. Call the list should fail. Hide Google on the collect page should fail. Unverified domain should draft and not send. Those failures are the product. A seat that never sees them will try them in a personal tab later.

If the new seat is a contractor, they need a role, not a forwarded password. Give them Scout and Forge before Relay. Give them the uuid. Do not give them DEPLOY. The agency and small-team pages are the longer role talk. The screen will still enforce the locks if the talk was sloppy.

How does a typed sentence stay email-first when the user asks for a phone blast?

A typed sentence stays email-first when the user asks for a phone blast because compliance intents are refused before classify picks a helpful send. Text the whole list, call these leads, and hide the one-star visitors are locks, not ambiguous mid-scores. Nova should say a discovered phone number is not consent, then point at email, the inbound line, or a form checkbox.

A search can still run if the first clause was a hunt. The SMS cannot. The chain stops with the lock named. Activity should show the refuse. A junior who then opens a personal SMS app next to Scout is off the product and still on the hook.

Voice on this screen is inbound. A demo number can be a Forge tool. A live line can be a deploy after they pay. There is no outbound cold path to slash-delegate into. If the sentence is have the robot call last year's leads, the answer is the same lock the Autonomy Matrix already painted red.

Client recaps after the sale do not grow a quiet mailbox either. Steward drafts. Relay sends. The agent screen does not become a retainer side door. Until you approve, or later write a standing rule that still passes the gate, the queue holds the artifact.