> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tied.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Training studio
> The front door, not a settings page. Where you shape the agent, argue with it in a sandbox, and publish a version you can roll back.
You sign up, open the Orchestration tab, and you are here. It is not buried three levels into account settings, and it is not something a solutions engineer opens on your behalf.
## The first screen is not empty
Because the [crawl already ran](/agents/knowledge) when your account was created, the studio opens on what Tied already learned about your business: the pages it found, the topics it extracted, and the gaps it noticed.
From there you choose how to proceed. Both paths are first-class.
A wizard. Confirm what was found, fill the gaps it flagged, set tone and escalation thresholds,
connect an app. Right if you want the shape of the thing laid out for you.
Start talking to the agent immediately and fix what is wrong as you find it. Right if you would
rather poke at it than be walked through it.
Either way, the first session ends with a working agent.
## Two modes, one surface
The deterministic flow builder and the reasoning agent live here together, clearly separated.
| Mode | What it is | What it does |
| - | - | - |
| **Deterministic** | Scripted flows and exact-match answers. Free. | Deflection and routing — catching the paths you can enumerate and putting everything else somewhere useful. |
| **Generative** | The reasoning agent. Paid. | Synthesises across sources, handles phrasing it has never seen, asks clarifying questions, decides when to escalate. |
**Every resolution belongs to the generative agent.** The free deterministic bot deflects and
routes; it does not claim to resolve anything. Keeping that line honest is what makes
[outcome pricing](/pricing/resolutions) defensible rather than a marketing claim — you are never
billed for a scripted flow matching a keyword.
The two hand off to each other. A flow that hits a question it does not handle passes the turn to the agent rather than straight to a human, and the flow keeps the structured paths it already handles.
## Who can open it
Admin-only is the **default**, not the ceiling.
You can grant studio access to an individual person, to everyone on a specific team, or by role permission — and you can grant *parts* of it, so someone can review knowledge without being able to publish a version. Every grant and every change is in the [audit log](/admin/audit-logs).
See [Custom roles](/admin/custom-roles).
## The sandbox
Talk to the agent exactly as a customer would, against your real knowledge and your real connected apps — **with writes disabled** unless you deliberately enable them against a test environment.
Every sandbox conversation is flagged as test traffic on the way out. It never bills, and it never enters your reported numbers.
The sandbox shows you the [trace](/agents/operations) as it runs: what the router decided, which specialist ran, what retrieval returned, which tool was called. When an answer is wrong, the trace usually tells you whether the fix is content, retrieval, or policy.
## Channel preview
Author once, then see the same conversation render on web chat, Apple Messages for Business and WhatsApp side by side.
**This does real work while you wait.** Apple Messages for Business onboarding runs four to six
weeks per brand. The preview lets you build everything now and see precisely what you are waiting
for, rather than discovering the shape of it on approval day.
Because [a flow never names a channel primitive](/agents/messages), what you preview is what ships — no re-authoring when the channel changes.
## Publishing is a version, not a save
Studio changes do not quietly mutate what is live. **Publishing creates a version**, and every version runs the evaluation suite before it goes anywhere.
The gate keys on **blast radius, not on what you pay**:
| Change | Gate |
| - | - |
| **Knowledge-only** | No hard block. The score is shown; you acknowledge it explicitly and publish. |
| **Anything that can write** to a connected system | The golden set must pass. No exceptions, on any plan. |
| **Enterprise accounts** | Always gated. |
**A write is a write whether you are on the cheapest plan or the largest contract.** A cheap plan
does not buy a looser gate on actions that change real state in your systems.
Rollback is one click, to any previous version.
This gate protects Tied as directly as it protects you: [negative signals automatically credit](/pricing/resolutions) your account, so a bad version costs us money. We would rather catch it in CI.
## The agent proposes back
The studio is also where the agent tells you what it learned.
It surfaces knowledge it inferred from watching your humans resolve things — questions it could not answer and a person then did, phrasings it did not recognise, flows it noticed being walked manually. You approve or reject each one.
Approved corrections write into the knowledge store with **provenance and a confidence flag** attached, so a fix persists and gets reused rather than being re-litigated one conversation at a time. Rejected ones stay rejected.
## Next
What the studio is editing.
What the publish gate actually runs.
The deterministic mode, in full.
Granting studio access without granting everything.
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.