Vault Agent decision pipeline

The Vault Agent’s Autoseek pipeline turns the current vault options and portfolio into target allocations, then hands required actions to the shared Execution Framework.

flowchart LR
  Options[Active vault options] --> Scope[Wallet execution checks]
  Scope --> Rules[Effective rule filters]
  Rules --> Scores[Model and rule-adjusted scores]
  Scores --> Targets[Capacity and portfolio utility]
  Targets --> Actions[Planned actions]
  Actions --> Execution[Execution Framework]
  Monitoring[Monitoring] -. new breach: eligible early .-> Autoseek[Autoseek run]
  Autoseek --> Options

Candidate discovery

Autoseek loads vault options for the agent’s base asset and removes inactive options. It also excludes experiment-blocked and permanently disabled vaults. The wallet’s execution path applies its own target checks, such as requiring a registered adapter or an approved address. A vault returned by the public vault-options API is not guaranteed to pass these checks.

Rules and eligibility

The Rule Manager supplies the effective Vault Agent rules: memory-derived user rules and computed defaults from the selected preset. explorative is the default preset; conservative is the other supported profile. The effective-rules API returns the user and default rules separately, and a user rule replaces a default of the same type. Autoseek also injects ignored-vault rules for blocked or non-depositable targets.

VaultRulesEngine applies exclusion and eligibility rules first: ignored vaults, vault-type filters, collateral exclusions, and configured liquidity floors. It then applies per-vault capacity and withdrawable-liquidity allocation limits. Forced allocations are considered only for eligible vaults and remain subject to their capacity, available balance, and the portfolio’s vault-count limit.

Scores and target allocations

The inference model scores options; an optional yield-curve rule can adjust those scores. APY alone does not determine the result. The engine evaluates candidate portfolios using adjusted scores, liquidity and concentration risk, and estimated transition costs, within the agent’s vault-count limit. It may keep a current allocation or choose a different target portfolio.

Monitoring and early runs

Monitoring is a separate feedback path, not an allocation or execution step. The Vault Monitor builds measurable rule boundaries for held vault positions and checks them continuously against live metrics. If a breach becomes triggered after the agent’s last Autoseek, the agent becomes eligible for its next run without waiting for the normal frequency limit. Monitoring does not choose a target or move funds; Autoseek re-evaluates the portfolio.

Execution handoff

Autoseek plans actions from the target allocation. The shared Execution Framework validates and submits only permitted protocol interactions. See Execution framework.

Explaining a decision

The agent actions API can return recorded action and decision details for supported runs. They describe the inputs and outcome for that run, not a promise about future allocations.