Agent layer
The agent layer is included for beta users. It runs on your own LLM key, so it can create model-provider costs — this page covers what agents do and how to keep those costs bounded.
What agents do
Hypothesis drafting. Given a market observation, the agent produces a testable strategy specification: entry, exit, sizing, universe.
Strategy authoring. The agent writes the Python implementation, registers it in the lab, and kicks off a screen run.
Debugging. If a strategy underperforms in the gauntlet, an agent inspects the trace and proposes parameter changes or structural edits.
Post-mortems. Every closed trade gets an automated post-mortem — why it triggered, why it exited, whether the outcome matches the thesis.
What agents do not do
- They do not place trades directly. All live orders go through the same promotion gate as human-authored strategies.
- They do not run unsupervised in the background without the daemon. If the daemon is off, agents are off.
- They do not skip the gauntlet. No matter how confident the agent sounds, the strategy graduates only if it passes walk-forward OOS.
Bring your own key
You supply the LLM key. Forven does not resell tokens. Supported providers:
- Anthropic (Claude Opus, Sonnet, Haiku)
- OpenAI (GPT family)
- Custom OpenAI-compatible endpoints
Keys live in the local keystore. They never leave your machine, never touch our servers.
Cost discipline
The agent layer can burn through tokens fast if you let it. Forven exposes hard caps per run and per day. The defaults are conservative; raise them only after you've watched how much a run actually costs.
Cost dashboards are still coarse during beta — treat them as a guide, not a precise ledger, and lean on the hard caps rather than after-the-fact reporting.