Trayntrayn.ai
RL environments for real-world AI agents

Trayn makes real work trainable

Record any workflow on any app. Train, evaluate, and verify your AI agents against hyper-realistic RL playgrounds.

The Problem

Why agents fail at real work

Real workflows span multiple apps and steps. Context gets lost across handoffs. UI changes, hidden state, and edge cases make agents fail silently.

Trayn captures those failure points and turns them into training data for the next run.

How Trayn Works

From real workflow to trained agent.

Capture. Train. Verify

RL Playgrounds and Training infrastructure

Captured workflows become practice environments. Agents run, get feedback, and improve over repeated reps.

app.trayn.ai/session/slack.com/c92d...
# announcements42 members
Sarah Chen████ ████2:38 PM

Forwarded the Acme Corp████ ████ contacts from Gmail████. Can someone update the CRM before the 3pm call?

Mike Johnson████ ███████2:40 PM

On it. Pulling the data into Salesforce██████████ now. @Sarah████ heads up — two of those emails bounced.

Acme Team████ TeamAPP2:41 PM

CRM sync: 3 contacts updated (sarah@acme.com████@████.com, j.doe@acme.com█.███@████.com, m.lee@acme.com█.███@████.com)

Lisa Park████ ████2:43 PM

Thanks Mike████! I'll prep the deck for Acme████ — meeting link is in the Google Cal████ ███ invite.

Message # announcements

Privacy-first capture

All personal data is detected and replaced offline in your browser before anything is stored. Names, emails, and org data are anonymized while preserving structural context.

$trayn --url https://app.trayn.ai/playground/gmail/abc123 --reps 3
Fetching task definition...
Launching browser environment...
Rep 1/3: Running agent against playground...
Grading step 1: click("compose") → correct
Grading step 2: fill("to", "jane@acme.com") → correct
Grading step 3: click("send") → correct
✓ Task completed — accuracy: 100%

SDK + CLI

Point the SDK at any playground URL. Your agent attempts the task, gets graded on each step, and stores memories for the next repetition.

app.trayn.ai/playground/gmail.com/a8f3...
Mail
Search mail
Primary
Social
Promotions
████ ██████
Re: Q4 ██████ review — Attached is the updated report for your...
2:34 PM
███ ██████
██████ onboarding — Welcome aboard! Here are the next steps...
2:28 PM
████████ ████
Notes — ████ sync — Action items: 1. ██████ to update the...
1:45 PM
██ ████████
Weekly ██████ digest — 12 tasks done, 3 pending review, 1 blocked...
12:10 PM
████ ███
Re: ██████ access — Approved. You should have access within...
11:02 AM
Grader
FAILED: click "Save Draft"
DO INSTEAD: click "Send" in toolbar
Stored for the next rep.

Training feedback loop

Agents practice real workflows in interactive playgrounds and get clear feedback on what to improve next.

memories.json
// Step grades from rep 1
{
"step": 3,
"action": "click('save-draft')",
- "outcome": "AVOID",
- "reason": "clicked wrong button"
+ "outcome": "REPEAT",
+ "correction": "click the Send button in toolbar"
+ "reason": "matched expected action"

Graded memories

Trayn records what worked and what failed, then uses those corrections in the next rep so agent behavior improves over time.

Start training your
agent today.

RL environments for real-world AI agents