Collect workflow data for
training continual learning models

Every finished run lands in the pool — question and answer included.
The problem

LLMs don't know your business

Off-the-shelf models miss what
makes your operation run.

They've never seen your ERP or your team's judgment calls — so they stall where the work gets specific.
Generic Assistant
Stalled 3
Invoice Reconciliation
Asked 4 clarifying questions
Vendor Onboarding
Wrong approval chain
Payroll Close
Missing policy context
Handed back 2
✗
Quarterly Board Deck
Returned to analyst
22m
✗
CRM Data Hygiene
Returned to ops
15m
Invoice Reconciliation
Reconcile March invoices in the ERP against vendor statements and flag any mismatches for review

Happy to help! I just need a few things first:

It needs to know:
  • Systems: Which ERP holds the invoices — and how does it get access?
  • Policy: What counts as a mismatch, and what tolerance applies?
  • Vendors: How do statement names map to your vendor records?
  • Process: Who reviews flags, and where do they land?

Every answer lives in your team's heads and tools — context a general-purpose model has never seen.

Describe your ERP setup…
? Waiting on you0 apps connected↑
The solution

Mya learns as it works

Traces become post-training
datasets for continual learning.

Every click, screen, and decision is captured and distilled into datasets — so your models keep learning how your business runs.
Mya Models
Capturing 2
Invoice Reconciliation
214 steps · 4 apps
Vendor Onboarding
96 steps · 2 apps
Training 1
finance-ops-v3
Epoch 2 of 4
Deployed 3
✓
finance-ops-v2
2,140 runs this week
v2
✓
crm-hygiene-v4
890 runs this week
v4
✓
deck-builder-v1
312 runs this week
v1
finance-ops-v3
Build a post-training dataset from this week's 12,400 reconciliation traces and continue training

Pipeline running end to end on your team's real work:

From traces to model:
  • Capture: Every click, screen, and decision recorded across your apps
  • Dataset: Traces de-identified and structured into 41k training examples
  • Post-train: The model learns your tools, policies, and edge cases
  • Evaluate: 94.2% end-to-end completion on held-out workflows

Every new run feeds the next dataset — the model keeps learning as your team works.

Run a workflow with this model…
◇ finance-ops-v3Staging ⌄↑
The data

Every run writes a trace.
Traces become datasets.

What Mya records while it works — the raw material of a post-training set.

Environment
What the user was looking at when they asked: the focused app, the app under the cursor, the browser tab, even the selected cells. Captured at submit, so every trace starts with real context.
Actions
Every command and tool call with its output, every edit — numbered steps, the same ones that tick by on the badge.
Screens
A screenshot of the target app on nearly every action — what the agent saw when it decided.
Judgment calls
When the agent needs a decision it stops and asks. The question and the answer stay in the trace — your team's judgment, made explicit.
Format
Plain JSON threads, grouped by app, streamed to disk as the run happens. Any run exports as readable text.

Try Mya now.