AI Chat is Catalio’s built-in conversational interface. Open it from the Ask bar at the top of any page, or on Home, and talk to one of your organization’s five agents. The agents can read your requirements catalog, personas, journeys, initiatives, and more. They can also create records, take you to a page, and pre-fill forms for you to review before saving.
Agents and Conversations
Every organization has the same five agents. Each answers in its own area:
- Product Manager shapes requirements, sorts priorities and writes acceptance criteria with you.
- Architect talks through designs, data models and the tradeoffs between them.
- QA Analyst designs test cases, finds coverage gaps and triages flaky tests with you.
- Security Analyst reviews auth and access, threat-models features and triages security findings.
- DevOps works through deployments, incidents and infrastructure config with you.
Software Engineer is no longer offered for new conversations, because every agent can now start sandbox tasks. Your existing Software Engineer conversations still open and work.
Every agent can read your catalog, but each one only creates or changes records in its own area. If you ask for something outside it, the agent can hand that part to the right agent.
You can hold many conversations with each agent. Pick an agent in the command center’s rail (the switcher at the top on smaller screens) to open its view: a conversation list, plus New conversation to start another. Opening an agent returns you to its newest conversation. On Home, Recent lists your recent conversations across agents. A new conversation starts untitled, and Catalio names it from its messages within a few minutes of your first exchange.
The open conversation has its own header. Click the pencil to rename it, or open the ⋯ menu to Pin it, Mark as unread, Rename, Export it as a Markdown file, or Delete it. Delete asks you to confirm first, because it removes the conversation and its messages for good. New starts another conversation with the same agent, and its ▾ lets you start one with any agent. On wide screens, an agent panel sits beside the conversation. On smaller screens and phones, choose Show agent panel (the panel icon at the top right of the conversation) to open it. Its Agent tab shows the agent’s status (Available, Working, or Needs you), Start new chat, View tasks, and Quick actions that start a new conversation with a prompt ready to finish. Its Thread tab shows when the conversation was created and updated, the tasks it started, and its tags and attachments.
Tag your conversations to keep them organized: + Add tag in the conversation header adds one of your own tags, and By tag in the conversation list’s filter menu shows only the conversations with that tag. Use the paperclip (Attach files) to add images, PDFs, and text files to a message, up to 25 MB each and 10 per message. Export lists each message’s attachments by name. When a reply includes code for a specific file, the code block shows as a file card with the file’s path, Copy, and line numbers.
On a phone, swipe a conversation left to show Mark read (or Mark unread), Pin, and Delete, or long-press it for Pin conversation, Mark as unread, Rename, Export, and Delete.
Your conversations are private. Nobody else in your organization can open them, org admins included.
Your Earlier Conversations
The first time you open Product Manager, your most recent conversation from before agents existed can become a Product Manager conversation, so that history carries on. Conversations tied to an initiative stay out of it.
Agent Profiles
Click an agent’s mark or name at the top of a conversation to open its profile: its name, label, and description, set by Catalio. The profile also has your Notifications toggle for that agent, which controls whether you are notified when one of its tasks finishes or needs your input. It is on until you turn it off.
Page Context
Each message also tells the agent which page you asked from. On a record’s page it knows which record “this” refers to, and on a create or edit form it can fill the form in for you (see below).
Messages
Every message comes from one of three sources:
- You — messages you type into the chat interface
- The agent — its responses
- Another agent — when an agent hands part of your ask to a teammate, the handoff and that agent’s answer show in your conversation, labeled with who answered
Handoffs stay within your own conversations. The Stop button ends them here: no more handoffs go out, and an answer still on its way won’t show in this conversation, though the other agent may finish it in its own conversation with you.
AI responses stream in real time: you see partial text arrive and a completion indicator when the response finishes. If something goes wrong, an error banner appears so you can retry.
While the AI works, a short status line shows what it is doing — for example, “Looking that up…” or “Searching requirements…” — so you have live progress without waiting for the full response.
Interactive Blocks
Sometimes the AI needs a specific answer before it can proceed. Instead of asking a free-form question, it presents a clickable interactive block:
- Choice — pick one option from a list (radio-style)
- Multi-select — pick multiple options (checkbox-style)
- Confirmation — yes or no decision
Click your answer directly in the conversation. The agent reads your selection and continues without you needing to type a reply.
What the AI Can Do
The AI is not limited to conversation. During a response it can take action in Catalio on your behalf. Tool calls are visible in the UI so you always know what the AI did.
Create and update catalog records
Creates, updates, or connects entities in your catalog — requirements, personas, journeys, jobs-to-be-done, use cases, policies, processes, capabilities, components, applications, and issue links.
Example triggers: “Create a requirement for single sign-on”, “Update the checkout persona’s description”, “Connect this requirement to the Payment initiative”
Manage initiatives
Creates or updates Initiatives — name, description, strategic intent, scope summary, process domains, and dates.
Example triggers: “Create a new initiative for the Q3 onboarding project”, “Update the initiative target date to September”
Navigate the app
Opens a specific page inside Catalio. The AI validates the destination before navigating — external URLs and admin paths are rejected.
Example triggers: “Show me the requirements list”, “Take me to the checkout journey”, “Go to initiative settings”
Pre-fill creation forms
When you already have a creation form open, the AI can suggest field values and wait for you to review and submit. Nothing is created until you do.
Ask for a new record from any other page and the AI creates it directly instead, then tells you what it made.
Example triggers (while on a /new page): “Draft a persona for finance managers”, “Fill this in for a CSV export requirement”
Pre-fill edit forms
When you already have an edit form open, the AI can propose changes for you to accept, adjust, or iterate on in chat before saving. Nothing is saved until you submit the form.
Ask for an edit from any other page and the AI updates the record directly instead, then tells you what changed.
Example triggers (while on a record’s edit page): “Change the title of this requirement”, “Update the priority to high”
Start tasks on your sandbox
When your organization has sandboxes turned on, any agent can start development work on your sandbox as a task: building, fixing or changing code, or running a command. The task appears as a card in the conversation with its status, and Open in sandbox opens its console. A task that is still running can show across your other agent conversations too. If you have more than one sandbox, the agent asks which one to use. The agent works out which repository to use from the records pinned to the conversation, and asks you when more than one matches. A task can work across several repositories. Each agent lists the tasks it started on its own Tasks tab.
Example triggers (to any agent): “Fix the failing login test on my sandbox”, “Run the migration script as a task”, “Build the next issue in the delivery queue”
Secrets and environment variables in your sandbox
A task builds and runs your code with the environment that the repository’s setup runbook gives it. Keep secrets in a secrets manager, never in a file on the sandbox:
- Sign in with a runbook step. Add a step to the runbook that signs in to your secrets manager, such as
doppler login,op signinfor 1Password,infisical login,aws sso login, orvault login, and mark it as needing you. The first time a task reaches that step, it stops with Setup needs you and the step’s name. Choose Open sandbox terminal, sign in once, then continue. - Load secrets at run time. Wrap the runbook’s build and run steps with your tool’s run command, such as
doppler run --,op run --, orinfisical run --. The values go straight into the process and are never written to disk. - Don’t use
.envfiles for secrets. Don’t write secrets into a.envfile on your sandbox, paste them into task instructions, or share them in a conversation. Agents can read and commit files in the checkout, and each task works in its own copy, so a file left in one task isn’t there for the next.
If your repository doesn’t use a secrets manager, Catalio can store its secrets instead. On the Application Settings page, the repository’s card has an Environment row. Choose Manage to add them. Secrets are write-only and masked in task output.
If a task’s install or build step fails, the agent names the step and offers Fix the runbook. Accepting starts a setup run from your current steps, and approving it saves the new steps. To check the saved steps without changing them, choose Re-validate on the repository’s runbook on the Application page. You need permission to edit the Application.
Ask structured questions
Presents an interactive question block when the AI needs structured input to continue.
Edit PRD sections
Applies section-scoped edits (add, update, append, remove) to an initiative’s PRD document, preserving untouched sections instead of regenerating the whole document.
Example triggers: “Add a Risks section to the PRD”, “Update the Success Metrics section”, “Remove the Open Questions section”
Search issue links
Searches issue links (GitHub, Linear, Azure DevOps, and manually linked Jira items) across your workspace by repository, milestone, labels, state, assignee, or free-text term.
Example triggers: “Find open GitHub issues for the billing repo”, “Search Linear issues labeled bug”
Suggest initiative groupings
Fetches loose (unlinked) issues and active initiatives for an application so the AI can propose how to group the issues into initiatives.
Example triggers: “Help me organize these loose issues into initiatives”, “Which initiative should this issue belong to?”
Discover catalog tools on demand
Instead of loading every catalog action upfront, the AI starts with a few list tools for quick overviews and counts (applications, initiatives, capabilities, features, requirements, and personas). For anything else it finds the right action with discover and runs it in the same turn. That covers read, search, count, and write actions across requirements, personas, policies, components, capabilities, use cases, journeys, processes, initiatives, test cases, and more. You see these as ordinary assistant actions in the chat — there is no separate tool picker to manage.
Multi-Tenant Isolation
Every conversation and message is scoped to your organization, and you can only open your own conversations. Background jobs that write the agents’ responses always run inside your organization.
Best Practices
Be specific about what you want. “Create a requirement for exporting user data as CSV, high priority, for the Finance persona” produces a better first draft than “add a data export requirement.”
Ask from the page you are working on. The agent knows which record you are looking at, so on an initiative’s page you can say “this initiative” without restating which one.
Pick the agent for the job. Ask Product Manager about requirements and priorities, Architect about designs and data models, QA Analyst about tests and coverage, Security Analyst about auth and threat models, and DevOps about deployments and incidents. Any of them can run work on your sandbox. Each conversation keeps its own history.
Iterate in chat before saving. When the AI pre-fills a form, review the values in the chat first. Say “change the description to…” and the form updates live. Save only when the draft looks right.
Retry on errors. If a response shows an error banner, use the retry button. Transient LLM failures are common and usually resolve on a second attempt. If a conversation’s history grows too large for the model, start a new conversation with that agent or ask in another agent’s conversation.
Ask questions freely. The AI can list records, explain fields, compare options, and walk through workflows. Use it as a research assistant, not just a form-filling shortcut.
Relationships at a Glance
| Entity | Relationship |
|---|---|
| Initiatives | Read and updated through the agents |
| Requirements | Created or updated through the agents |
| Journeys | Created or updated through the agents |
| AI Skills | Each agent answers every message in its conversations with its skill |
Next Steps
- Explore AI Skills to see the expertise behind each agent
- Learn about Document Extraction to see how uploaded files generate Change Proposals for human review
- See how Change Proposals provide the human-in-the-loop review workflow for AI-generated suggestions
Support
If conversations stop responding or errors persist after retrying, check the LLM provider configuration at Settings > LLM Providers. The AI uses the provider configured for your organization. See Supported AI Providers for the options, and AI Features and Data Privacy for how your data is handled when the AI processes it.
