EvoTrader: a self-evolving multi-agent trading system on ADK (open source, research) #7307
AdamCYH
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
I built a research project on ADK and open-sourced it this week: a team of agents that trades a stock or ETF through your own Robinhood account inside hard limits, plus an evolution agent that reads the scored record of past decisions and the source code, then proposes strategy and code changes for a human to approve. The hypothesis — that an agent can re-tune an algorithm and propose strategies as market conditions change — is unproven, and the project publishes no trading results. It runs in practice mode by default.
Repo: https://github.com/AdamCYH/evotrader (Apache 2.0). Architecture notes:
docs/architecture.md.ADK-specific things that might save someone time:
transfer_to_agentis one-way. A customBaseAgentreached that way cannot hand control back the way anLlmAgentsub-agent can (its function call returns). We moved the strategy role to a custom agent and silently lost the risk-manager and executor stages for a run of cycles — every cycle "held", and looked healthy. An automated review caught it by noticing no order had been possible. If your pipeline depends on a role returning, reach it as a tool call, not a transfer.FilteredMcpToolsetwrapper does it.BaseAgent, with a pluggable backend registry so other CLIs could slot in. It withholds the API key from the subprocess so usage can only bill the subscription, and reads the SDK's rate-limit events to fall back to the API path when quota runs low.Happy to answer questions about any of it, and genuinely interested in how others handle the one-way transfer problem.
All reactions