Agent Governance & Alpha Factory
Coming SoonRoute /dashboard/agents · Flag enableAgentAssist
Agents in Podium substitute for headcount at the research and operations edges — never for the irreversible decision in the middle. Every agent is a governed automation: it reads freely, it proposes drafts to a human gate, and it never mutates the ledger autonomously. The alpha factory applies the same discipline to systematic signal discovery.
Read / propose / mutate
Agent capability is classified by a hard boundary. Read access is free. Propose actions produce drafts that land in the cockpit approval inbox. Mutate actions — anything that changes the live path or the ledger — are never agent-autonomous; they are human-gated or deterministic-engine-only.
Role agents
Role agents assist specific functions (for example data, research, and operations) and are each bound to a deterministic verifier so their output is checkable rather than trusted. Each agent is itself governed, with an evaluation harness tracked in the evidence vault.
Alpha factory
The alpha factory runs an agentic propose → evaluate → select loop over the alpha-operator DSL the SDK already exposes. It is paired with a mandatory trial-count-aware overfitting gate (deflated Sharpe, probability of backtest overfitting): without that gate, automated expression search is a confidence-inflation machine. Promising candidates still pass through signal governance and human approval before going live.
Related
- AI-Native Capability Layer — the contract agents call
- Signal Governance — where candidates get promoted
- Alpha Operators (SDK) — the DSL the factory searches over