Creating an agent

You make an agent on one page: its name and instructions, when it runs, and the tools it may use, with everything else under Advanced settings already filled in. You can also describe the agent in a chat and let Strata draft it for you; see Create an agent from a chat.

Start from a template, or from scratch

Open Agents and click Create agent. The split button offers:

• Create a custom agent — opens the form with empty defaults.
• Create from a template — opens the Templates view. Pick a template and the form opens prefilled with that template's name, instructions, schedule, tone, tools, and (for org templates) model and run limits. You can edit anything before creating it.

If you launch from an org template marked locked (the admin disabled customization), there is no form — clicking the template deploys the agent in one step with the admin's exact configuration. See /docs/agents/templates for the locked-vs-editable distinction.

The form

The page is headed Create a new agent (Edit and the agent's name when you edit one). From top to bottom:

1. Name, Description and Instructions, with the model the agent runs on in the instructions' footer.
2. Schedule — when it runs.
3. Tools — what it may use.
4. Advanced settings, closed — tone, response, email and whether it's active, each already set.

Cancel and Create agent (Save changes when you edit) stay at the bottom of the screen while you scroll, so you can create the agent from anywhere on the page. Everything you leave alone keeps its default: a new agent runs Weekdays only at 09:00 in your timezone, writes in the Professional tone, uses the Standard response size, sends no email, and is active once created.

Name and instructions

1. Name (required, 120 characters max) — e.g. Daily morning briefing. The line under it checks availability as you type: Name is available or You already have an agent with that name. Names are unique per user.
2. Description (optional, 500 characters max) — one line about what this agent does.
3. Instructions (required, 8,000 characters max) — what the agent does on every run. The line under the box says why to be explicit: The agent can’t ask you anything while it runs, so say what to do, what to use and what to send back. The placeholder reads:

Scan unread client emails. Group them by urgency (reply today / this week / FYI) and save the brief to OneDrive Strata/Briefings as morning-brief.md.

The instructions box's footer holds:

• The model the agent runs on, a button showing the model's logo and name. Its list is drawn live from your organization's Model Catalog, so you see exactly the models your admin has enabled — nothing more — with older generations in a Legacy models group at the end (still supported, not recommended for new agents). When a model retires, agents pinned to it move to its replacement on the retirement date, unless your organization turned automatic migration off. If an agent is pinned to a model your organization no longer offers, it stays in the list so you can see it and switch it. For the full live catalog and how models are governed, see Models.
• An info icon (See model token costs) that opens the AI model costs modal, listing each enabled model with its provider and live per-million-token input and output rates.
• A counter of the instructions' length out of 8,000.

Note

The model list shows the models your organization has enabled. The model an agent is actually allowed to run on is the intersection of what's enabled for your organization, your role (including any role bound to one of your groups), and any per-user model access your admin set for you. Strata checks that intersection when you create an agent or switch it to another model: a model outside it is refused with <Model> isn't available to you: your role doesn't allow it. Pick one of: …. (or your organization doesn't allow it / it isn't in your own model list). It checks again when you resume an agent and before every run. A save that keeps the model and doesn't resume the agent isn't checked, so you can still pause, rename or reschedule an agent whose model you've lost. If Strata can't read your access when you create an agent, switch models or resume, the change goes through and the run decides. An agent always runs on its pinned model. Your organization can't turn that model off while an active agent is pinned to it (Strata refuses until those agents are moved to another model; paused agents don't count), but an admin can still narrow your role, a role one of your groups gives you, or your own model access so it leaves the model out; the agent's next run then fails with a message naming the setting that leaves it out. If that run is its schedule, its webhook or another agent invoking it, Strata also pauses the agent; Run now is refused without pausing it, and the page shows Could not start run with the reason. Switch it to an allowed model, or have an admin allow the model again, then press Resume (or turn Active back on as you save). Resume on a model you still can't use is refused: This agent can't resume. … Edit it to use another model, then resume it.

Schedule

Pick a frequency, a time, and a timezone. The options in Frequency are:

• Manual only (no schedule) — agent only runs when you click Run now or call its webhook. The schedule sentence reads Your agent has no schedule. It only runs when you click Run now or call its webhook.
• Every N minutes — pick any cadence from 1 to 59 minutes. A platform floor of 5 minutes always applies (and your org's minimum-interval rule on top of it), so a cadence under 5 minutes — or under whatever longer interval your admin set — is rejected on save.
• Every 15 minutes, Every 30 minutes, Every hour, Every 6 hours — fixed cadences.
• Every day — pick the hour and minute.
• Weekdays only — Mon–Fri at the picked time.
• Specific days — pick weekdays plus hour and minute (defaults to Mon–Fri).
• Custom (advanced) — type a raw five-field cron expression.

Underneath the friendly controls, an expandable Use a raw cron expression disclosure lets you override any friendly selection with a raw cron. Below that, the schedule sentence updates live to summarize what you picked (e.g. "Your agent will run weekdays at 9:00 am in America/New York with no end date."). Changing the frequency redraws the schedule's fields; anything you typed elsewhere on the page stays.

Click Add an expiration date to set a date after which the agent stops running automatically. Type a YYYY-MM-DD value or pick one from the calendar. Backspace clears it.

See /docs/agents/schedules for cron details, minimum-interval enforcement, and how the next-run time is computed.

Tools

Turn on only what this agent needs. Tools that send email or write files act without asking you during a run, so turn them on only when you trust the instructions. The tools are grouped, one line per group showing how many of its tools are on (0 of 4); click a group to open it and see its switches. A group with a tool on starts open, so what the agent may do is on the page without a click. The section's own count reads No tools on, or N of M on.

• Database — Run SQL queries (the run_sql tool).
• Outlook — Search inbox, Read full emails, Read attachments, Send email (the Send email switch is flagged Sensitive because agent runs bypass the chat-time confirmation prompt).
• OneDrive — Search files, Read files, Save files (the Save files switch is flagged Sensitive).
• Jira — Search issues (the search_jira tool: issues, service requests, and service-desk queues) and Create issues and comment (the write_jira tool, flagged Sensitive — an agent run has nobody to show an approval prompt to, and it only works if your organization allows Jira writes).
• Confluence — Search pages (the search_confluence tool) and Create pages and comment (the write_confluence tool, flagged Sensitive, and likewise gated on your organization allowing Confluence writes).
• Generation — Generate Excel, Generate PDF.
• Code & composition — Analysis (sandboxed), Invoke another agent (calls another agent you own as a sub-task; max nesting depth 3; a paused agent is refused rather than run), Sub-agents (splits a run into up to 5 parallel, read-only workers).
• Web — Search the web (live lookups for current events, prices, and documentation via Brave Search).

The Web, Jira and Confluence groups are hidden when that extension isn't allowed for you, rather than shown and refused later, so your form may show fewer groups than the list above. The form uses the same extension check Strata applies before each run: your role, widened by any role bound to one of your groups, narrowed by your organization's extension list in Ceiling mode, minus any extension an admin turned off for you. A role saved with no extensions ticked allows none. The Allowed agent tools list an admin can set for your agents isn't reflected here, so a group can still show whose tools that list strips before the run. Strata reads your access when the app loads and again whenever you come back to its tab or window (at most once a minute), so an admin's change to your role, groups or extensions reaches the form without a reload: it shows the next time the form is drawn. If Strata couldn't read your access in full the last time it checked, the form hides none of these groups rather than guess; the check before each run still decides. The other groups always show; a tool whose extension you aren't allowed is stripped before each run instead. Jira and Confluence are two separate groups over one Atlassian connection — an admin can allow one and deny the other, and only the allowed one appears.

A recommendation. When you leave the Instructions box with new instructions in it, Strata asks the AI for the smallest set of tools the job needs. While it works, the Tools section says Analyzing your task to recommend the right tools…; then an AI recommendation lists them with Apply recommendations and Not now. Apply recommendations turns those tools on and opens their groups (it doesn't turn any tool off); Not now dismisses it. Instructions the form opened with, unchanged, ask for nothing (an agent you're editing already has its tools), and neither does an agent started from a template, which comes with its own tools.

See /docs/agents/permissions for a deeper dive on safe defaults, the dry-run sandbox, and admin-level controls.

Advanced settings

Advanced settings is closed when the form opens. The line under it says what is set inside (for a new agent, Professional tone · Standard responses · No email), so you can leave it closed without missing a choice. Click it to open:

• Tone — how the agent writes. Pick one of five tones; the sample under the tabs shows how it reads:

• Response — Response mode (Auto, Technical, or Simple; the same modes as chat) and Response size, three preset tiles:
  • Quick — short answers (60s runtime, 5,000 tokens).
  • Standard — recommended for most agents (300s runtime, 20,000 tokens).
  • Thorough — deep analysis or long reports (600s runtime, 60,000 tokens).
  • Expand Advanced — set the exact numbers to type runtime (30–600 seconds) and max tokens (500–100,000) directly.
• Notifications — every run's results land in your Strata inbox. Email me when this agent runs (off by default) also sends a plain email each time it runs; turned on, it shows:
  • When to email — After every run or Only on failures.
  • Send to (optional) — leave it blank to send to your account email, or type any address.
  • Include the run output in the email — off → just the status; on → the agent's output (truncated to 12,000 characters) is included.
• Active once created — on by default. Turn it off to create the agent paused (Turn off to create it paused. You can resume it any time from the Agents list.). When you edit an existing agent, the same switch reads Active (Turn off to pause it.).

See /docs/agents/notifications for the system sender and deliverability tips.

Click Create agent

Create agent (or Save changes when editing) checks the whole form from the top. If something is missing or wrong, the first such field is marked and gets the focus, and a toast titled Please complete all required fields says what to fix; when it's under Advanced settings (an email address in Send to, say), that section opens to show it. Otherwise Strata POSTs to /api/agents. On success the new agent's own page opens, with a toast titled Agent created saying when it runs (<name> is scheduled: Weekdays at 9:00 AM, America/New York., or <name> runs when you press Run now or call its webhook. without a schedule). Until it has run once, a Try it now note on its page offers a Test run. When editing, the toast reads Agent updated / Changes saved. and you're back where you started the edit: the agent's page or the list.

Cancel and Back leave the form. If you changed anything, they ask first: Discard your changes? — Nothing you changed here has been saved. — with Discard and Keep editing.

Create an agent from a chat

You don't have to open the form. In the Strata web app, ask for the agent in the normal chat box, for example "Every Monday at 9am, summarize my open Jira issues and email me the list." Strata turns your words into a draft:

• a name and the instructions the agent follows on every run (written to stand on their own, because a run can't read your chat);
• a schedule in your own timezone, with the next run, or Only when you run it;
• the tools it needs, and no more;
• where results go: always your Strata inbox, plus an email after every run or only when a run fails if you asked for one;
• the model it runs on.

The draft appears as a card under Strata's answer, and the answer's action trail shows Drafted an agent. Nothing is created until you choose Create agent. The card says so at the bottom.

• Create agent creates the agent exactly as the card shows it. After it's created, the card reads Agent created with a line such as Scheduled. Next run Mon, Sep 21 at 9:00 AM, America/New York time. and an Open agent link that opens it on the Agents page.
• Open in Agents page opens the form filled in with the draft, so you can change anything before you choose Create agent.
• To change the draft, tell Strata what to change ("make it 8am", "also check Confluence"). A new card appears with the change; create the one you want.

Connectors the agent needs

Once you create the agent, its card lists the connectors its tools use (Outlook, OneDrive, SQL Server, Jira or Confluence), each marked Connected or with a Connect button that runs the usual connection. While one isn't connected, the card says the agent can't use it until you connect it, and offers Don't use a connector.

Choosing Don't use a connector says so in the chat. Strata then drafts the change: the agent without that connector's tools, its instructions rewritten so they don't rely on it, no tool it didn't have unless the job needs one, and a sentence saying what the agent can no longer do. That card reads Changes to an agent and offers Update agent; nothing changes until you choose it, and the card then reads Agent updated.

If the draft uses a tool that acts without asking on every run, such as Send email, Save files or Write to Jira, the card shows an orange line naming it (Acts without asking on each run: Send email.) and its Instructions section starts open, so you can see who the agent emails and what it changes before you create it.

In a saved chat the card stays with the answer, so it is still there after a reload, and a created card still shows Agent created. If you regenerate the answer, each version keeps its own cards.

The same checks as the form

A draft is checked against everything the Agents page checks, before the card is shown and again when you choose Create agent: whether your role, your account and your organization allow you to create agents, your agent quota, trial limits, the 5-minute schedule floor and your organization's minimum interval, and that the name is not one you already use. A draft can also only use models and tools you are allowed to give an agent.

When something stands in the way, Strata tells you why in plain words instead of showing a card. If the agent needs a connector you haven't connected, Strata still drafts it; once you create it, the card recommends connecting it (see Connectors the agent needs). If the check fails only when you choose Create agent (for example, you reached your agent quota in the meantime), the card shows the reason and Create agent stays available to try again.

Only in the web app

Draft cards and Create agent exist only in the Strata web app. Ask for an agent in the Outlook add-in, Teams, the mobile app or the API, and Strata answers that agents are created in the web app, points you to its Agents page, and offers to write out the name, schedule and instructions so you can paste them in.

What can go wrong

• Name and task prompt are required. — the name or the instructions are empty. Fill in both.
• You already have an agent with that name. Pick a different name. — names are unique per user (case-insensitive). The inline pill catches this while you type; if the name was taken between typing and clicking Create agent, you'll see the toast version of the same message.
• the model list has not loaded yet, try again in a moment — Strata checks the agent's model against the live Model Catalog, and it hadn't finished loading (usually just after an update). Wait a few seconds and save again.
• <Model> isn't available to you: your role doesn't allow it. Pick one of: ... — the picker lists every model your organization has enabled, but neither your role nor any role one of your groups gives you allows this one. The reason can also be your organization doesn't allow it or it isn't in your own model list (a per-user limit your admin set). Pick one of the models the message lists, or ask your admin to allow it. Switching an agent to another model in the form is checked the same way. A save that leaves the model unchanged isn't, so you can still pause, rename or reschedule an agent whose model you've since lost; turning Active back on is refused (This agent can't resume. …) until you switch it to a model you can use.
• Pick at least one day of the week. — you chose Specific days but didn't select any.
• A cron expression is required. — you switched to Custom (advanced) but left the cron field blank.
• Invalid cron expression: ... — the cron string didn't parse. Use five fields (minute hour day-of-month month day-of-week), e.g. 0 9 * * 1-5.
• Agents are disabled for your role or organization (or Agents are disabled for your organization) — your role's canCreateAgents is off, or an admin has turned on Freeze agent changes (lock_agents) for the organization. Contact your admin.
• Agent creation has been disabled for your account by your administrator. — your admin set the per-user lock. You can still pause/resume and star existing agents; you can't create new ones or edit existing prompts/schedules until the lock is removed. See /docs/agents/per-user-governance.
• Agent quota reached: your organization allows up to N agent(s) per user. Delete an existing agent to make room. — you've hit your per-user max. Delete an unused agent and retry.
• Agent quota reached: your organization currently blocks new agents. Contact your administrator. — your admin set Agents per member to 0.
• Schedule is too frequent: ... — your schedule fires more often than the allowed minimum. Two forms of this message exist: ...scheduled agents run at most once every 5 minutes... is the platform floor that applies to every organization, and ...your organization requires at least Ns between runs... is a longer interval your admin set. Pick a less frequent cadence.

The Agents page shows these exact sentences when you create or edit an agent, and a chat draft card shows the same sentence when Create agent is refused.

Related

• /docs/agents — Agents overview.
• /docs/agents/schedules — schedules in detail.
• /docs/agents/notifications — notifications in detail.
• /docs/agents/webhooks — webhook triggers.
• /docs/agents/permissions — capabilities and safety.
• /docs/agents/templates — templates.
• /docs/chat/models — the Model Catalog and how models are governed.