Skip to main content

Agent Treasury Example

Updated August 2026 · Live on testnet as an SDK example.

An honest framing first: this is a policy loop, not an AI. No LLM, no model, no autonomy claims — a deterministic program (packages/sdk/examples/agent-treasury.ts) that demonstrates how an automated treasury would use Leontief.

The loop​

An agent earning USDC income runs four rules on a cycle:

  1. Earn — USDC revenue arrives at the treasury address.
  2. Wrap — idle USDC above a working-capital floor is converted to the RWA and deposited; the treasury now holds yield-bearing ld-shares instead of idle cash.
  3. Borrow, don't sell — when the treasury needs liquidity, it pledges shares and borrows USDC against them (inside a health-factor floor), keeping the yield position intact.
  4. Repay — incoming revenue pays debt down first when the health factor tightens.

Onboarding (trustlines + SEP-8 issuer authorization) is deliberately outside the loop — it is a compliance step the issuer performs, not something an autonomous treasury can do for itself (scripts/agent_onboard.sh).

The "public SDK only" rule​

The example imports nothing but the exported @leontief/sdk surface. If the loop ever needs a private import, that is an SDK gap to close first — this rule is what forced tokenBalance, transfer, and maxBorrowForHealthFactor into the public API. The example doubles as the SDK's dogfood test.

Guardrails​

  • Health-factor floor enforced before every borrow (maxBorrowForHealthFactor, floor-rounded in the user-facing direction).
  • Working-capital floor: the loop never wraps the last unit of operating cash.
  • Every action simulates first and aborts on any contract error; typed errors, no retries past policy bounds.
  • Testnet only; the example is documentation, not a product.

Where this goes next: the same pattern under Autopilot's on-chain policy constraints (Tranche 2 scope) — the loop's authority bounded by the ledger instead of by its own code.