How the agent actually works.

Tokenized assets pay income: stocks pay dividends, property funds pay rent, credit pays interest. Collecting that income is manual, easy to miss and easy to fumble. Dividend Garden is an agent that watches for those events, proves they really happened, and moves the proceeds where you decided in advance.

It is non-custodial. Your assets stay in your own wallet or smart account, and the agent holds a narrow, revocable permission over generated income - never over your position.

• You keep the assets. The agent never takes possession of them.
• You set the rules once: how income splits, and the limits it must respect.
• The agent acts only on income events, only for allowlisted assets.
• Every action publishes an on-chain receipt you can verify yourself.
• When anything looks wrong, nothing executes. That is a NO-GO.

Every routed dollar goes through the same six stages. The stages are fixed code, not a judgement call - the same event with the same plan always produces the same outcome.

1. 1 · Detect

   An indexer watches your allowlisted assets for income events - a dividend multiplier update on a stock token, a distribution posted by a property fund.

2. 2 · Verify

   The event is checked against an independent source: the corporate action feed for stocks, the issuer distribution proof for funds. Unverified means nothing happens.

3. 3 · Resolve

   The agent computes exactly how much income was generated - the newly created portion only, never your principal or market appreciation.

4. 4 · Decide

   The amount is checked against your plan: minimum payout, monthly limit, allowlists, slippage. Any breach ends the event as a NO-GO.

5. 5 · Execute

   Inside your limits, the agent claims and routes the income into Reinvest, Earn and Reserve using pre-approved contracts and destinations.

6. 6 · Publish

   A receipt goes on-chain: the source event, the verification, the gross, the fee, the split, the timestamps.

The plan is the whole contract between you and the agent. It has two halves: where income goes, and what the agent is never allowed to exceed.

Income splits across three routes - Reinvest buys more of the asset that paid, Earn moves proceeds into a yield destination, Reserve holds stablecoins. The three always total 100%.

Changes apply to the next event. Nothing already executed is retroactively re-routed.

Rather than handing the agent a key to everything, you sign a scope. The scope lives in an ERC-4337 smart account and delegates a session key that can only make calls matching your rules.

• Scoped to income: the session key cannot transfer or sell your position.
• Scoped to targets: every contract the agent may call is pinned in advance.
• Scoped to destinations: funds move only to addresses your plan names.
• Scoped in time: the session key expires and has to be renewed deliberately.
• Revocable: one transaction ends the agent’s authority entirely.

A revoke does not need the agent to cooperate. It is your account, and the permission is yours to withdraw.

A NO-GO is the agent declining to act. It is a successful protective outcome, not a failure - the safest thing an automated system can do with an event it cannot stand behind is nothing at all.

• Oracle unavailable - the data source did not answer, so the agent waited instead of guessing.
• Event not verified - the on-chain event did not match the independent source.
• Slippage too high - the quote drifted past your limit.
• Amount below minimum - not worth the fees, by your own rule.
• Monthly limit reached - your ceiling, respected.

A NO-GO leaves your funds exactly where they were, records the reason in your feed, and charges no fee. The agent retries when conditions are valid again.

Automation you cannot audit is just trust with extra steps. Every executed event publishes a receipt with enough detail for anyone - not only you - to confirm the agent followed your plan.

Each asset class produces income differently, so each has its own verification path. Unsupported assets cannot be added at all - the allowlist is enforced, not advisory.

Dividend Harvest deserves a note. When a tokenized stock reinvests a dividend, your share count does not change - the token’s value multiplier rises instead. The agent measures that rise, converts only the newly generated portion, and routes it. Your position is never sold.

Availability depends on the issuer and your jurisdiction, and is shown per asset before you add it.

A performance fee on routed income only - 0.25%. There is no fee on your principal, no management fee, and no fee when an event ends in NO-GO. If the agent does not route anything, it does not earn anything.

The parts that decide are deterministic. The parts that guess are advisory and cannot execute.

1. Deterministic core

   Eligibility, amounts, limit checks and routing are plain code with no model in the path. Same inputs, same decision, every time - and reproducible after the fact from the receipt.

2. Advisory LLM

   The language model explains events and drafts summaries. It has no authority to approve, size or execute anything. A model failure degrades the explanation, never the money.

3. Adapters

   Chains, oracles, issuers and execution venues sit behind interfaces, so a swapped provider does not reach into the core.

4. Simulation before execution

   Every route is simulated against live state first. A simulation that fails or drifts past your limits becomes a NO-GO instead of a transaction.

The controls below are enforced on-chain by the permission scope, not by the agent choosing to behave.

• The agent cannot touch your principal - Permissions cover generated income only. Your position stays yours.
• Funds go only where you allow - Unknown addresses are rejected. Destinations come from your plan.
• Allowlisted assets only - The agent watches the assets you approved, and nothing else.
• Allowlisted execution contracts - Every call target is pinned in advance.
• Monthly execution limit - A hard ceiling you set. Reaching it stops execution, not your income.
• Slippage limit - A bad quote produces a NO-GO instead of a bad fill.
• Emergency pause - Pause the agent or revoke permissions at any time, in one transaction.
• Full on-chain receipts - What happened, why, how much, where it went - all verifiable.

Pause and revoke are always available to you, in one transaction, without the agent’s involvement.