Multi-Agent Pipeline

A full pipeline instead of a single bot: one agent collects data, one trains models on a GPU, and one executes trades from its own wallet.

Create three agents#

Fund your deposit in Settings > Funds so all three agents have runway. Give sol_executor initial wallet funding and turn on Agent pays for itself so its runtime costs come from its own wallet.

Set up the collector#

Collect SOL market data, order book summaries, and derived features every minute.
Write raw data to disk and maintain a cleaned dataset for training.
Also expose a tiny status page on port 7000 that shows whether ingestion is healthy.

Set up the trainer#

Read the collector's cleaned datasets, retrain the signal model every night,
save the latest checkpoint and evaluation report, and publish the newest model
artifact to a GitHub repo or object-storage location that the executor can fetch.

Set up the executor#

Load the latest approved model, generate live SOL trade signals,
pull the current model artifact from the shared repo or storage location,
execute only when confidence clears a threshold,
use the agent wallet for trades,
and expose a dashboard on port 8080 with wallet balance, open positions,
daily PnL, and model version currently in production.

Monitor#

Expose ports:

• 7000 on sol_collector for ingestion health
• 8080 on sol_executor for live trading visibility

Use both dashboards together: the collector page to check data freshness, the executor page to check whether the current model is profitable.

Iterate without rebuilding#

When something drifts, send a targeted prompt instead of starting over:

The live strategy is overtrading during volatile periods.
Retrain with volatility-aware position sizing, deploy the new checkpoint,
and add a panel to the dashboard that compares today's model against yesterday's.

If the executor's strategy starts making money, its wallet can keep paying for its own runtime without pulling from your main deposit balance.