Studio
Studio is where you build and run agents — autonomous AI workers that do recurring work for you, such as monitoring a source, summarizing activity, drafting content, or syncing data across your connected tools. Studio is the home screen of CatalEx: when you sign in, you land here.
An agent has its own instructions, a set of tools it can use, success metrics (KPIs), an optional set of run schedules, and a memory that grows as it works. You can chat with an agent to test it, run it on demand, or let it run automatically on a schedule.
The Studio home page
Your agents appear as a grid of cards. Studio is personal — you see the agents you created (agent limits, however, are shared across your whole company). Across the top you have:
- Search agents… — filter by name, description, or tag.
- Sort — Most recent, Most run, or Alphabetical.
- New agent — start building. If your company is at its agent limit, this instead points you to upgrade.
- Tag filters — pills for Design, Engineering, Ops, Sales, Marketing, Research, HR, Product, Finance, Personal. Select one or more to narrow the grid.
If you have no agents yet, the page invites you to Build your first agent with a prompt box and a row of starter templates.
The agent card
Each card shows the agent's avatar, name, and description, plus a quick-glance status row:
- A status pill — Paused or Draft (an active agent shows no pill). Cards needing attention (a run Running or Approval needed) float to the top of the grid.
- The version (e.g. v1), run count (e.g. 12 runs), and N scheduled when the agent has recurring schedules.
- The time and outcome of the last run.
- Two small icons next to the name jump straight to the agent's Impact (chart) and Evolution (DNA) tabs, and a row of tool icons shows the integrations it's bound to. A
!badge flags a tool that isn't connected yet.
From the card you can Run now (or Resume if paused), Pause an active agent, or Delete it. Clicking the card opens the agent's full detail view.
Creating an agent
Click New agent. You can start in three ways:
- Describe it in plain English. Type what you want — for example, "Track competitor pricing pages and alert me when anything changes." This opens a short conversation where an assistant interviews you to understand the job.
- Start from a template. Pick a ready-made blueprint such as Product Manager, Account Strategist, Code Reviewer, Feedback Synthesizer, or Content Creator.
- Browse the full catalog. Open Agent Templates to search ready-made agents by category.
The build conversation
The assistant asks questions to refine the agent's purpose, the tools it needs, and how success should be measured. Answer in your own words. Once you've exchanged enough detail (at least a few back-and-forth turns), the Create Agent button becomes available.
When you click it, CatalEx synthesizes a proposed agent — Analyzing conversation → Discovering tools → Crafting instructions → Proposing KPIs — and hands you a review screen.
An agent draft only exists during this session. If you leave agent creation before finishing, CatalEx warns you that the draft will be lost.
Review and create
On the review screen, everything is editable before the agent is created:
| Section | What you set |
|---|---|
| Name | The agent's name (click the pencil to rename). |
| Description | A short summary of what it does. |
| Category | The domain the agent belongs to — helps you find and group agents. |
| Instructions | The full operating instructions, in Markdown. |
| Tools | The integrations the agent may use. Add or remove any. |
| Schedules | Optional recurring runs (see Scheduling). |
| KPIs | The success metrics it is judged against (up to 8). |
Click Create Agent to finish. The new agent appears on your Studio home page.
You don't have to get everything perfect at creation. Instructions, tools, schedules, KPIs, and memory can all be edited later from the agent's detail view.
Working with an agent
Opening an agent reveals a detail panel with eight tabs. The header always shows the agent's name, a clickable status badge (Active / Paused / Draft), a Schedules button, and a Run now button, plus a delete control.
Chat
Chat is the default tab and the best way to test an agent. Type a message, and the agent responds — streaming its reasoning, calling tools, and asking for your approval when it wants to take a sensitive action. Use the ＋ button to start a new conversation.
If the agent proposes an action that requires approval, you can Approve, Reject, or tell it to do something else in your own words.
Files produced during a chat. When a run finishes having created files, the panel widens and splits in two — your chat stays on the left, and a session artifacts pane opens on the right with a tab per file over the built-in viewer described under Artifacts. Each file has the same quick actions (open in a new tab, download, share). Minimise the pane to a » N pill at the edge of the chat, and click it to bring the files back. The pane opens on its own only when a live run produces new files; reopening a finished chat shows a compact summary you can expand. (On mobile, files appear in the Artifacts tab instead of a split view.)
Overview
The agent's core Instructions (editable Markdown) and its Domains (tags). This is where you refine what the agent knows how to do.
Memory
What the agent has learned and remembers across conversations. Each memory is a short Fact (something to remember) or Rule (something to always do, e.g. "Always cite sources"), up to 600 characters. You can add, edit, or delete memories, pin the important ones so they're never auto-replaced, or clear all. The agent also picks up memories on its own as it works.
Tools
Manage which integrations the agent can use. Your connected tools appear under Agent's tools; browse Other Available Tools to add more. For each tool you can choose which connected accounts the agent may act through, and set per-action guardrails. See Tools for how connecting and guardrails work.
Runs
The history of the agent's runs — both scheduled runs and the ones you trigger with Run now. Each row shows the run's title, status, and when it was last updated. Statuses include Starting, Running, Waiting for input (an approval is needed), Completed, Failed, and Cancelled. Click any run to read its full transcript. Filter by status or date, or search across runs.
Artifacts
Files the agent saved during its runs — reports, documents, spreadsheets, pages, and the like. Search them and switch between tile and list views. Click any file to open it in a built-in viewer that adapts to the format:
- PDFs, images, Markdown, code, and CSV (as a table) render inline.
- HTML renders in a sandboxed preview, with a view source option.
- Very large or unrecognised files offer a download instead of a preview.
Each file has quick actions: open in a new tab, Download, and Share. Sharing creates a link that anyone can open in a browser without a CatalEx login — choose how long it lives (1 day, 1 week, 1 month, or Forever), copy it, and revoke it whenever you like. The same viewer also appears beside the Chat tab whenever a run produces files.
Impact
Measures the agent against its KPIs. After each run, CatalEx scores the agent (0–100) on each KPI and records what worked and what needs improvement. The tab shows current scores, trends, and sparklines over your chosen window (Last 7 days / Last 30 days), and lets you add or remove KPIs. Pausing measurement here also pauses evolution.
How are KPIs measured?
KPI measurement is self-contained: after a run finishes, an automated evaluator reviews that run and nothing else — the request, the agent's responses, every tool call with its result (success or error), and execution stats such as duration, turns, and tokens. It then scores each approved KPI from 0 to 100 with a short reasoning, plus what worked well and what needs improvement.
Think of it as a teaching assistant grading a student's submitted work: the evaluator can judge the quality of the submission itself, but it can't follow the student home to see whether the advice paid off. It has no access to your external systems, analytics, or anything that happens after the run ends.
That means a KPI is only as good as what the conversation can show:
| Measurable (visible in the run) | Not measurable (outside the run) |
|---|---|
| Completeness — did the response cover everything the request asked for? | Business outcomes — revenue, conversions, CSAT, retention |
| Grounding — are claims and figures backed by data the agent actually fetched? | Events in external systems after the run (e.g. whether a fix held in production) |
| Follow-through — does every claimed action have a successful tool call behind it? | Rates aggregated over events the run never saw (e.g. "% of tickets deflected") |
| Structure and actionability of the deliverable | Wall-clock business timelines measured in days or weeks |
| Efficiency — focused execution without wasted turns or tool errors | Whether people adopted the agent's recommendations |
Phrase KPIs about the work product the agent delivers in the conversation — "every flagged discrepancy names the account, amount, and suspected cause" — rather than downstream outcomes like "reconciliation error rate below 2%". Each run gets its own score, so trends across runs build up automatically on this tab.
Evolution
This is where the agent improves itself. Based on its KPI observations, the agent proposes changes to its instructions, memory, tools, and KPIs. You review each proposal — Approve, Edit, Compare (a side-by-side diff), or Reject — then the approved changes become a new version. Every version is kept in a history timeline, so you can compare versions or restore an earlier one at any time. Use Evolve now to trigger a cycle on demand. Evolution requires impact measurement to be turned on.
Scheduling recurring runs
Agents can run automatically on a recurring basis. Schedules are managed from the Schedules button in the agent's header (it reads Create schedule when you have none, or Schedules (N) once you do).
An agent can have any number of schedules. To add one, click Add schedule and choose:
| Field | Options |
|---|---|
| Frequency | Daily, Hourly, or Weekly |
| Time | The local time of day (Daily and Weekly) |
| Every | The interval in hours (Hourly) |
| Days | One or more weekdays (Weekly) |
| Prompt | (optional) what the agent should do on each run |
The prompt is what makes schedules powerful: each schedule can carry its own instructions, so one agent can have several schedules that each kick off a different recurring job (for example, a morning digest and a Friday report). Leave it blank to run the agent with its default instructions.
Each saved schedule shows its cadence in plain English in your local timezone. When a schedule fires, the run appears in the Runs tab just like any other run, where you can open it and read what the agent did.
Times you enter and see are always in your local timezone.
Pausing an agent
Pausing an agent stops its scheduled runs while preserving everything — its instructions, tools, memories, and history. Pause it from the card, the header status badge, or Run now → Resume. Resume it whenever you like.
FAQ
What's the difference between chatting with an agent and running it? Chat is an interactive conversation you drive in real time — ideal for testing. A run is the agent working on its own, either because you clicked Run now or because a schedule fired. Runs appear in the Runs tab.
Do I need to be technical to build an agent? No. You describe what you want in plain English and review the result. The technical surfaces — custom scripts and custom MCP servers — are optional and for advanced users.
Why is an agent asking me to approve an action? Sensitive tool actions require approval by default (a "guardrail"). You can change which actions run automatically — see Tools.
What does "Evolution" actually change? With your approval, it updates the agent's instructions, memory, tools, and KPIs to improve its scores — then saves the result as a new version you can restore.
Can I share an agent with a teammate? Agents themselves aren't shared between people — each is owned by the person who created it. You can share individual artifacts an agent produces via a link (see the Artifacts tab).
Can I pause an agent without deleting it? Yes. Pausing an agent stops its scheduled runs while preserving everything. Resume it whenever you like.