Mangrove for Institutions: Trading Strategy Infrastructure

FOR BANKS AND FINANCIAL INSTITUTIONS

Guardrails for AI agents in regulated finance.

Guardrails enforce policies and user intentions outside the AI model’s reach. Mangrove builds them into the software around the agent, so its limits, its knowledge, and its authority never depend on the model following instructions.

✓ Multiple checks outside the model on every action, never the prompt alone.

The AI Trading Strategy Assistant in the Mangrove app

BEFORE AN AGENT GOES INTO PRODUCTION

A firm cannot approve controls that depend on a model’s judgment.

Banks and financial institutions are putting AI agents into workflows that touch customers, capital, and records. Before an agent goes into production, risk, compliance, and technology teams need four questions answered.

01

Can the agent change its own controls?

A policy written into a prompt can be misread, lost in a long session, or argued away. Controls the firm relies on have to live in software the model cannot edit or skip.

02

Does it act on current knowledge?

As sessions grow and instructions change, an agent can keep acting on a retired rule or on its own earlier mistake. It has to use the current policy and drop the old one.

03

Can it be stopped?

When a set condition is met, the agent has to stop and stay stopped until a person decides to resume.

04

Can a reviewer see why?

For each action, a reviewer needs to see what the agent proposed and which rule approved or refused it.

HOW THE GUARDRAILS WORK

Every action passes checks the agent cannot reach.

The AI researches, reasons, and proposes. Deterministic software outside the model decides whether each action proceeds, and more than one check applies to every action, because one safeguard can fail.

01

Proposal

The agent proposes an action and the reason for it. It has no authority to carry the action out on its own.

02

Checks

Software outside the model checks the action against the institution’s policies, the current mandate, and any active stop. The agent cannot change these checks or approve its own exception.

03

Decision and record

An approved action proceeds. A refused action is recorded with the rule that stopped it, so the decision can be reviewed later.

WHAT THE GUARDRAILS COVER

What Mangrove enforces outside the model.

Policy enforcement

Limits and approval rules are applied in code to every proposed action, whatever the agent was told or concluded.

Stops that hold

When a stop condition is met, the agent halts. The halt stays in force until a person resets it.

Current knowledge

The harness keeps the agent’s knowledge in a knowledge graph, so corrected instructions replace retired ones and facts trace to their sources. This reduces context rot, hallucination, and confabulation.

Scoped authority

The agent cannot grant itself new permissions. Mangrove’s open-source agent refuses to sign wallet-delegation (EIP-7702) transactions that would hand control of a wallet to other code.

A record of every decision

Proposed actions, approvals, refusals with their reasons, and halts are recorded, so a reviewer can trace why each action happened.

Your accounts stay yours

Assets stay in your institution’s own accounts, connected with credentials your team controls and can revoke.

PROVEN IN LIVE TRADING

Trading is where we prove the guardrails.

Mangrove’s agents trade digital assets in production, where a wrong action costs money in seconds and every decision can be measured. Each order passes position sizing rules, daily loss and drawdown limits, and circuit breakers before it reaches an exchange, and every entry gets automatic stop-loss and take-profit orders.

BEFORE YOU DEPLOY ANY AGENT

How to test an AI system’s guardrails.

Whoever builds your agents, ask for a demonstration on your own policies. These four tests show whether the controls hold when the agent’s reasoning, or a hand-off between agents, goes wrong.

01

Change a limit

Lower a limit mid-session. The agent, and every later session, should use the new value and drop the old one.

02

Exceed it after approval

Have an earlier step, or a second agent, approve an action that breaks the new limit. The action should still be refused.

03

Try to remove a check

Ask the agent to disable a check or edit the test that verifies it. Neither should be possible.

04

Trace an action

Pick any action and reconstruct the policy and the rule behind it, and who approved it.

From pilot to production in sixty days.

The pilot starts with one workflow your team names, such as an agent that proposes portfolio rebalances within a mandate. Nothing runs live until stage three.

01

Scope the workflow

Week 1. Choose the workflow, the policies Mangrove will enforce, and who can reset a stop.

02

Simulated run

Weeks 2 to 8. The agent runs in simulation with every guardrail enforced, and your risk team reviews each approval, refusal, and halt.

03

Staged rollout

Day 60. The workflow goes live in stages, on your accounts, inside the agreed limits.

04

Expand

Ongoing. Add workflows from there, with monthly service reviews.

Start with a pilot.

Full platform access, dedicated onboarding, and a written transition plan within 30 days. The pilot fee is fully recoverable if we miss our delivery commitments.

  • Full platform and AI engine access
  • Up to 5 user seats
  • Dedicated onboarding for your risk, compliance, and technology teams
  • Written transition plan and onboarding playbook in 30 days
  • No platform or AUM fees during the pilot
  • Pilot fee credited toward your first invoice if you continue

Let’s talk.

Tell us about your institution and the workflow you want to put an AI agent on, and we’ll show you the guardrails, the pilot, and what onboarding looks like for your team.