ANODA Discuss your product

Enterprise AI agent builder, permissions in sight

Nexus is Larridin’s agentic AI product for company teams.We designed its Agent Builder, Computer Use and Nexus Vault, embedded in Larridin’s team.

Nexus Agent Builder: the conversation that set up a daily briefing agent on the left, its flow from the daily trigger through two actions on the right
200+
screens for the Nexus scope on desktop web
3
product areas Agent Builder, Computer Use and Nexus Vault
3
person team embedded in Larridin’s product team

Nexus is Larridin’s agentic AI product for people at large companies, built around one idea: describe a job, and an agent runs it on a schedule across the company’s apps. We joined Larridin’s product team as embedded AI product designers and designed the Agent Builder, Computer Use and Nexus Vault for desktop web, with a design system and UI kit for the new agent patterns.

What an agent goes through in Nexus

One chain, from picking an agent to handing it a login:

  1. Library

    • Templates
    • Popular in the company
    • My agents

  2. Builder

    • Conversation
    • Manual editing
    • Test run

  3. Trigger

    • Schedule
    • Repeats
    • Manual start

  4. Actions

    • Prompt and model
    • Connectors
    • Order

  5. Runs

    • History
    • Status
    • Back into chat

  6. Browser tasks

    • Plan
    • Live activity
    • Take over

  7. Vault

    • Personal details
    • Credentials
    • Cards
Conversation Precision
Chat to set it up; a visible flow and plain fields to fix the details.
Autonomy Oversight
The agent works alone; people see each step and can stop it or step in.
Speed Consent
Access asked for inside the task, item by item, not in a settings maze first.

One agent, read at a glance

Found
An agent is a trigger, a chain of actions, a status and a set of connected apps, and a page that lists them as equal fields doesn’t show what the agent does, or why it won’t run.
Decided
The agent page reads in the order of a run: what the agent is and how it starts, its flow from the start point through each action, then its details and a warning when a connector still needs signing in.
For the user
Whoever opens the agent, the person who built it or a colleague checking it, sees what it does and what blocks it without opening the editor.
Nexus agent page for Daily Morning Briefing: name, description and a Scheduled badge, the flow from the daily trigger through a knowledge search, an email summary and a Slack message, the Agent Details card and an Authenticate Connectors warning
After: the page of a scheduled briefing agent; names and dates are example data

  1. What it is

    Name, purpose and how it runs, with its status, team and connectors beside.

  2. How it starts, what’s missing

    The trigger at the top of the flow, and a warning while a connector waits for sign-in.

  3. What it does, in order

    Each action with its app and model, in the order it runs.

How the work progressed

Flows first, then structure in grey, then the interface and the system under it, and a handoff for each feature, all inside Larridin’s team.

  1. User Flow

    How an agent gets from a sentence to a scheduled run, and where a person must confirm or grant access

    Artifact User flows and screen maps for the builder, browser tasks and the Vault

  2. Wireframes

    Where the conversation ends and the editor takes over, and how a browser task shows what it’s doing

    Artifact Wireframes of the builder, the agent page and the task view

  3. UI

    How the agent patterns fit the product people already used for chat

    Artifact Desktop screens and states, and a design system and UI kit for the Nexus scope

  4. Delivery

    What engineering needed to build each flow without guessing

    Artifact A clickable prototype and developer handoff

Three ways to hand work to an agent

Found
People come to agents from different places: from scratch, from a chat that just gave a good answer, or with a one-off job for a website, and a single way in would push two of them down the wrong path.
Decided
Each way in got its own path to the same agent: a guided conversation, a builder panel opened beside the chat, and a Computer Use switch in the message box.

  1. From a blank page

    Describe the agent or pick a template; the assistant then asks what it needs, starting with the trigger.

    Open the full map
    Three steps of the guided start: the empty builder with templates, the request typed in, and the assistant’s first question about the trigger

  2. From a chat that worked

    The Agents button opens a builder beside the conversation, so a good answer becomes an agent that runs again.

    Open the full map
    Three steps from a chat to an agent: a chat answer with documents and Slack threads, the Create New Agent panel beside it, and the Add New Action dialog filled in

  3. A job in the browser

    Switch on Computer Use in the message box and describe the job; the assistant restates the plan and says where it will pause.

    Open the full map
    Three steps of a browser task: the empty chat, the Computer Use switch with its tooltip, and the assistant confirming a room booking plan while it sets up the browser
Two Nexus menus: the compact rail with New Chat, Prompt Library, Projects and Agents, and the full menu with Agents, the Admin area with Analytics and Settings, and recent chats

A rail and a full menu

Left: the compact rail around the builder and browser tasks. Right: the full menu, with the admin area and recent chats. Prompt Library and Analytics belong to AI Gateway, Larridin’s earlier product, not to this case.

  • User flows
  • Screen maps
  • Navigation states
  • Clickable prototype

Six jobs, from the first sentence to the password

The builder first, then the agents a team shares, then jobs in the browser and the data they may use. Every name, date, address and card on these screens is example data.

An agent set up by answering questions

Found
A blank form with a trigger, a prompt, a model and connectors asks people to think like the system before they know what the agent should do.
Decided
The assistant turns the request into a short interview: one card per decision, answers picked from chips or typed, each finished step marked done, and a way to skip straight to the manual builder.
Agent Creation conversation: the first action marked completed, then a card asking which email connector sends the briefing, with Gmail picked, and who receives it

A trigger people can read back

Found
“Daily” means little without a start date, a time and a rule for repeats, and schedule syntax is not something a team lead should have to write.
Decided
A start-point panel: scheduled or started by hand, a date, a time and repeats in plain words, echoed on the flow as “Daily at 10:15 AM”.
Manual builder with the Start Point: Trigger panel: Scheduled selected, start date, 10:15 AM and the frequency list open on Daily

Actions in the order they run

Found
An agent’s steps depend on each other, and a list of settings hides which step feeds the next.
Decided
Each action is a card on a flow under the start point; its prompt, model and connector sit in a side panel, and the cards change order by dragging.
Manual builder with the Search in company knowledge action lifted and dragged on the flow, and the Action #1 panel with its prompt, model and connector

Agents a team can find and run again

Found
A useful agent built by one person stays invisible to the next, who builds the same thing again, and a failed run leaves no trace.
Decided
A library that leads with the organisation’s most popular agents, filters by team and connector, and marks each agent as scheduled or manual; a history tab lists every run with its status and a way back into chat.
Agents library with Suited For and Connectors filters, a Popular in Your Organization carousel, and Marketing and Finance agents with Manual or Scheduled badges
Agents History table with agent name, run type, start time, status badges for In Progress, Failed, Canceled and Completed, and Try in Chat or Start Agent actions

A browser job you can watch

Found
An agent working through a website out of sight is exactly the black box a company won’t trust with a booking or a payment.
Decided
The job runs in a live browser view inside the chat, with the agent’s steps listed beside it and Stop, Hide activity and Take over browser always at hand; before it starts, the agent says where it will pause for approval.
Computer Use in the chat: a live browser on a workspace booking site with a map of New York, the agent’s steps on the right, and Stop, Hide Activity and Take Over Browser above
Take-over window: You are controlling Nexus’ browser, with a Finish Controlling button over the booking site

A vault instead of pasted passwords

Found
Browser jobs need names, logins and card details, and typing them into a chat puts them where they don’t belong.
Decided
Nexus Vault keeps personal details, credentials and cards on one settings page, and a task gets only the items the person connects to it.
Settings, Vault tab: general and additional information, work and home addresses, credentials for Google, Jira, Slack, Salesforce and WeWork, and the Personal Data Vault note

Agent patterns in the product’s own language

Found
Agents brought patterns the product didn’t have yet: flows, run states and access requests. Drawn anew for each feature, they would have made three products.
Decided
The Nexus patterns became components on the product’s existing foundations, reused wherever they appear: one flow card, one status badge, one consent sheet.

One consent sheet for everything an agent may touch

Apps an agent connects to and Vault items a browser job may use are asked for the same way: what, why, a check per item, then one confirm.

  • Authorize Apps to Use the Agent: Gmail and Slack connected, then what to expect, with Cancel and Finish Setup
    Agent Builder: Apps the agent needs, each signed in separately.
  • Connect Your Vault: personal information, credentials and credit card, each with its own selection, then Manage in Vault, Cancel and Connect
    Computer Use: Vault items shared with one task.

An action card, from hover to its new place

The card that holds each step: hovered, with its drag handle, lifted, and over its new place with a slot opening for it.

  • Action card Search in company knowledge, hovered, with its handle and remove button
    Hovered
  • The same action card with its drag handle under the pointer
    Handle
  • The action card lifted and dragged over the flow
    Lifted
  • The action card carried below the email action, a dashed slot marking where it will land
    Over its new place

Where the product ended up

Nexus got a complete design for its agentic scope: agents set up in conversation or by hand, triggers and actions on a visible flow, a shared library and run history, browser jobs people can watch, stop or take over, and a Vault that shares data item by item.

It came with user flows, wireframes, screen maps, a clickable prototype, a design system and UI kit, and developer handoff. Engineering wasn’t ours, and this page shows the design, not what shipped.

screens
200+
product areas: Agent Builder, Computer Use, Nexus Vault
3
roles: people who run agents, admins and reviewers
3
person team, embedded in Larridin’s product team
3

In the client's words

The design work gave the agent-building experience a clearer structure. We appreciated the way agent configuration, actions and permissions were brought into a set of screens that could be reviewed before implementation.

Nexus Team Nexus

5.0

Afraid a chat-built agent hides the settings that matter?

In Nexus, we designed agents that start as a conversation and end as a flow people can read back: the trigger, every action in order, each app it may touch, all editable by hand. Bring us your agent builder, and we’ll show where the chat should stop and the controls take over.

Discuss your product Explore more case studies

What we brought to Nexus

The product

The interface and the work