Treasury Copilot is an Infrastructure built for autonomous agents without giving them unrestricted access to human funds. currently on testnet
Abuja, Nigeria
Joined August 2026
- Tweets47
- Following28
- Followers8
- Likes185
Pinned Tweet
Treasury Copilot is a payment control platform for autonomous AI agents.
AI agents needs to pay for APIs, cloud services and other resources.
But giving an agent a private key is dangerous
We are building on @GenLayer for AI Agent Payments
Stop giving your Agents Funds
Give them a Policy and a Spending Delegation
@GenLayer x @MetaMask x @arc x @1shotapi x @circle made this possible
Let Your Agent Move Your USDC Based on your POLICY
"Moving money is one problem. Deciding what should happen when the transaction goes wrong is another."
Proud to see @kstellana among the three experts in @beincrypto's State of AI Agent Payments 2026.
beincrypto.com/uncomfortable…
We are a policy controlled payment system for AI agents.
- Humans delegate limited spending permissions through MetaMask.
- Agents request payments through an API
- GenLayer verifies each request against an onchain policy,
- 1Shot executes only finalized approvals.
Crypto theft exceeded $3.4 billion in 2025.
Chainalysis found that attacks against centralized services increasingly targeted private key infrastructure and signing processes.
Now AI agents are becoming economic actors.
Giving them financial authority without clear limits expands the risk.
(chainalysis.com (chainalysis.com/blog/crypto-…))
The problem is not simply that agents can make payments.
The problem is deciding:
→ Which agent is requesting the payment
→ What it is allowed to purchase
→ How much it can spend
→ Which recipients it can pay
→ When its authority expires
→ How that authority can be revoked
This is why we built Treasury Copilot.
[@TreasuryAIPilot ]
Treasury Copilot is the financial control layer for autonomous AI agents.
It lets agents request stablecoin payments through an API while owners retain custody and control.
Built with @GenLayer.
The owner registers an agent and defines its financial mandate:
→ Spending limits
→ Approved recipients
→ Permitted purposes
→ Policy conditions
→ Bounded wallet authority
The agent can request a payment.
It cannot silently expand its own authority.
Every request passes through layered controls.
First, deterministic rules check exact conditions such as the amount, asset, recipient, agent identity, and policy binding.
Then @GenLayer reviews whether the purpose and context of the request match the owner’s policy.
This distinction matters.
A valid signature only proves that a key signed a transaction.
It does not prove that the transaction was appropriate.
It does not explain why the agent paid that recipient, whether the purchase matched its role, or whether the amount respected its mandate.
@owasp describes Excessive Agency as a vulnerability caused by excessive functionality, permissions, or autonomy.
Prompt injection can turn that excessive authority into damaging actions.
For a financial agent, the damage may be an irreversible payment. (genai.owasp.org
(genai.owasp.org/llmrisk/llm0…))
Treasury Copilot is designed around a simple principle:
The model must never define its own financial limits.
The owner defines the boundary.
Deterministic controls enforce the boundary.
Contextual review evaluates the request inside that boundary.
This is not only our view.
@a16z has identified unresolved infrastructure questions around agent payments:
→ Who is the agent acting for?
→ What is its allowable purchasing scope?
→ Which merchants can it use?
→ How much is it authorized to spend?
There is still no agreed standard for assigning and verifying that scope. (a16z.com (a16z.com/newsletter/agent-pa…))
@NIST is also working on standards for AI agent identity and authorization.
Its research highlights identification, authorization, auditing, accountability, and controls against prompt injection as necessary parts of trusted agent adoption. (nist.gov (nist.gov/news-events/news/20…))
The identity problem is already larger than most organizations realize.
@CyberArk reports that financial services teams manage 96 machine identities for every human identity.
That does not mean there are 96 AI agents per employee.
It means organizations already face massive nonhuman identity sprawl before autonomous agents reach full scale. (cyberark.com
(cyberark.com/resources/finan…))
The usage curve is moving quickly.
@GoldmanSachs Research expects agent driven token consumption to grow 24 times between 2026 and 2030 as consumer and enterprise adoption expands.
More agent activity means more actions, more purchases, and more financial decisions that require enforceable controls. (goldmansachs.com
(goldmansachs.com/insights/ar…))
Merchants need controls too.
@Visa says merchants have historically treated automated traffic as bots and often blocked it.
As agentic commerce grows, merchants need a way to distinguish authorized commercial agents from malicious automation. (developer.visa.com
(developer.visa.com/capabilit…))
Wallet security is necessary, but it is not the entire system.
@MetaMask recommends treating an agent as an unfamiliar transaction proposer rather than a trusted signer.
Keys should remain isolated.
Permissions should be scoped and revocable.
Transactions should be checked before signing. (metamask.io (metamask.io/news/agentic-wal…))
Treasury Copilot adds the business policy and accountability layer around that wallet authority.
Wallets handle custody and signing.
Payment protocols handle settlement.
Treasury Copilot governs why an agent is allowed to spend and records what happened.
The problems Treasury Copilot addresses:
→ Agents operating beyond their mandate
→ Broad and persistent financial permissions
→ Prompt injection reaching payment authority
→ Uncontrolled agent budgets
→ Missing transaction context
→ Weak revocation and recovery
→ Incomplete audit history
Organizations deploying agents need more than hope.
They need:
→ Identity per agent
→ Authority per role
→ Budget per workflow
→ Policy per treasury
→ Evidence per decision
→ Revocation when something changes
→ A complete record after execution
The goal is not to remove humans from financial control.
The goal is to let humans define the rules once, let agents operate inside those rules, and bring people back in when a request exceeds the mandate.
That is bounded autonomy.
You do not manually manage network packets whenever you open a website.
Agent developers should not need to rebuild wallet permissions, payment queues, policy checks, retries, revocation, and audit infrastructure for every agent.
These controls should be infrastructure.
Treasury Copilot is currently a working testnet MVP.
treasurycopilot.app/
The current product includes:
→ @GenLayer StudioNet policy review
→ @MetaMask bounded permissions
→ @base Sepolia test USDC execution
→ Prompt injection denial testing
→ Idempotent payment handling
→ Auditable request history
Ten active testers have used the product, including a GenLayer Labs cofounder.
Sixty six automated tests pass across the web application and EVM contracts.
This is technical validation, not production scale, revenue, endorsement, or commercial partnership.
The agent economy will not scale on unrestricted authority.
It will scale on identity, policy, bounded permissions, verifiable intent, revocation, and evidence.
Give agents a budget.
Give them a mandate.
Do not give them unrestricted control.
Treasury Copilot is building that control layer.
We are Live on @arc
We integrated CCTP so you can deposit on Arc and Recieve USDC on any other Chain
Well we have been iterating and finding new ways to scale
Look forward to our integration of @circle
An agent wallet answers: "Can this key sign?"
Treasury Copilot answers:
- Should this payment be allowed?
- Is it within the agent's policy?
- What evidence supports it?
- Why did it execute or stop?
Treasury Copilot retweeted
Everyone building agent commerce is solving the happy path.
Payment rails, identity, how Agent A finds Agent B. That part is getting crowded fast.
Payments are getting solved.
Identity is getting solved.
Interoperability is getting solved.
What happens when two agents disagree about whether a contract was actually fulfilled is not solved, and that's the part that actually costs money.
Nobody is building for the moment the deal breaks.
I've been writing @GenLayer intelligent contracts since early 2026, and the same pattern keeps showing up:
you can encode a rule, but you can't encode reasonable.
A refund policy, a delivery confirmation, a dispute over quality of work, these need judgment, not just execution.
Traditional smart contracts execute perfectly and judge nothing.
That's fine until the two parties don't agree on what happened.
An agent to agent Ajudication layer is needed.
Agent A hands off a task, Agent B delivers the work, and something has to decide if what came back actually matches what was asked for.
That decision is the whole product. Not a footnote.
Think about why this is the expensive part.
A failed payment is a solved problem, retry it or refund it.
A failed handoff between two AI agents is not solved, because did this agent actually do what I asked is a judgment call, not a status code.
You can't retry your way out of a disagreement about quality.
If nearly $9 trillion in agent transactions is coming by 2030, every one of those is a potential dispute with no judge.
's answer is to not trust a single model's opinion on it.
GenLayer's answer isn't one model deciding.
Multiple validators reason over the same delivered work independently, and consensus only requires they land on the same meaning, not the same wording.
That distinction matters more than it sounds.
A single AI adjudicator becomes a single point of failure the moment people learn how it behaves.
That's the difference between one AI grading a submission and several examiners agreeing it's actually correct.
A validator set that only needs to converge on meaning is much harder to game, because gaming it means predicting how several independent reasoning processes land on the same judgment, not just tricking one.
By 2030 agents are projected to move close to $9 trillion.
Every one of those transactions can end in "that's not what we agreed."
Almost nobody is building for that moment right now. That's the gap worth building in.
A Court Case takes months or years to finalize and agent transacts in seconds and minutes
Its not nice to make an agent wait that long just to resole it's Conflicts
Well now you know why @GenLayer Matters
By 2030, AI agents will move nearly $9 trillion.
Every one of those transactions can end in a disagreement, and almost nobody is preparing for it.
That's why we built GenLayer, the adjudication layer for contracts that can actually think and start judging what's fair.
AI agents should be able to buy the resources they need.
They should not need your treasury private key to do it.
Treasury Copilot gives agents a scoped payment API while owners keep custody,
set limits, and audit every outcome.
Cracked Founder
Sudo He Is Dave
Crypto theft exceeded $3.4 billion in 2025.
Chainalysis found that attacks against centralized services increasingly targeted private key infrastructure and signing processes.
Now AI agents are becoming economic actors.
Giving them financial authority without clear limits expands the risk.
(chainalysis.com (chainalysis.com/blog/crypto-…))
The problem is not simply that agents can make payments.
The problem is deciding:
→ Which agent is requesting the payment
→ What it is allowed to purchase
→ How much it can spend
→ Which recipients it can pay
→ When its authority expires
→ How that authority can be revoked
This is why we built Treasury Copilot.
[@TreasuryAIPilot ]
Treasury Copilot is the financial control layer for autonomous AI agents.
It lets agents request stablecoin payments through an API while owners retain custody and control.
Built with @GenLayer.
The owner registers an agent and defines its financial mandate:
→ Spending limits
→ Approved recipients
→ Permitted purposes
→ Policy conditions
→ Bounded wallet authority
The agent can request a payment.
It cannot silently expand its own authority.
Every request passes through layered controls.
First, deterministic rules check exact conditions such as the amount, asset, recipient, agent identity, and policy binding.
Then @GenLayer reviews whether the purpose and context of the request match the owner’s policy.
This distinction matters.
A valid signature only proves that a key signed a transaction.
It does not prove that the transaction was appropriate.
It does not explain why the agent paid that recipient, whether the purchase matched its role, or whether the amount respected its mandate.
@owasp describes Excessive Agency as a vulnerability caused by excessive functionality, permissions, or autonomy.
Prompt injection can turn that excessive authority into damaging actions.
For a financial agent, the damage may be an irreversible payment. (genai.owasp.org
(genai.owasp.org/llmrisk/llm0…))
Treasury Copilot is designed around a simple principle:
The model must never define its own financial limits.
The owner defines the boundary.
Deterministic controls enforce the boundary.
Contextual review evaluates the request inside that boundary.
This is not only our view.
@a16z has identified unresolved infrastructure questions around agent payments:
→ Who is the agent acting for?
→ What is its allowable purchasing scope?
→ Which merchants can it use?
→ How much is it authorized to spend?
There is still no agreed standard for assigning and verifying that scope. (a16z.com (a16z.com/newsletter/agent-pa…))
@NIST is also working on standards for AI agent identity and authorization.
Its research highlights identification, authorization, auditing, accountability, and controls against prompt injection as necessary parts of trusted agent adoption. (nist.gov (nist.gov/news-events/news/20…))
The identity problem is already larger than most organizations realize.
@CyberArk reports that financial services teams manage 96 machine identities for every human identity.
That does not mean there are 96 AI agents per employee.
It means organizations already face massive nonhuman identity sprawl before autonomous agents reach full scale. (cyberark.com
(cyberark.com/resources/finan…))
The usage curve is moving quickly.
@GoldmanSachs Research expects agent driven token consumption to grow 24 times between 2026 and 2030 as consumer and enterprise adoption expands.
More agent activity means more actions, more purchases, and more financial decisions that require enforceable controls. (goldmansachs.com
(goldmansachs.com/insights/ar…))
Merchants need controls too.
@Visa says merchants have historically treated automated traffic as bots and often blocked it.
As agentic commerce grows, merchants need a way to distinguish authorized commercial agents from malicious automation. (developer.visa.com
(developer.visa.com/capabilit…))
Wallet security is necessary, but it is not the entire system.
@MetaMask recommends treating an agent as an unfamiliar transaction proposer rather than a trusted signer.
Keys should remain isolated.
Permissions should be scoped and revocable.
Transactions should be checked before signing. (metamask.io (metamask.io/news/agentic-wal…))
Treasury Copilot adds the business policy and accountability layer around that wallet authority.
Wallets handle custody and signing.
Payment protocols handle settlement.
Treasury Copilot governs why an agent is allowed to spend and records what happened.
The problems Treasury Copilot addresses:
→ Agents operating beyond their mandate
→ Broad and persistent financial permissions
→ Prompt injection reaching payment authority
→ Uncontrolled agent budgets
→ Missing transaction context
→ Weak revocation and recovery
→ Incomplete audit history
Organizations deploying agents need more than hope.
They need:
→ Identity per agent
→ Authority per role
→ Budget per workflow
→ Policy per treasury
→ Evidence per decision
→ Revocation when something changes
→ A complete record after execution
The goal is not to remove humans from financial control.
The goal is to let humans define the rules once, let agents operate inside those rules, and bring people back in when a request exceeds the mandate.
That is bounded autonomy.
You do not manually manage network packets whenever you open a website.
Agent developers should not need to rebuild wallet permissions, payment queues, policy checks, retries, revocation, and audit infrastructure for every agent.
These controls should be infrastructure.
Treasury Copilot is currently a working testnet MVP.
treasurycopilot.app/
The current product includes:
→ @GenLayer StudioNet policy review
→ @MetaMask bounded permissions
→ @base Sepolia test USDC execution
→ Prompt injection denial testing
→ Idempotent payment handling
→ Auditable request history
Ten active testers have used the product, including a GenLayer Labs cofounder.
Sixty six automated tests pass across the web application and EVM contracts.
This is technical validation, not production scale, revenue, endorsement, or commercial partnership.
The agent economy will not scale on unrestricted authority.
It will scale on identity, policy, bounded permissions, verifiable intent, revocation, and evidence.
Give agents a budget.
Give them a mandate.
Do not give them unrestricted control.
Treasury Copilot is building that control layer.
Treasury Copilot retweeted
Crypto theft exceeded $3.4 billion in 2025.
Chainalysis found that attacks against centralized services increasingly targeted private key infrastructure and signing processes.
Now AI agents are becoming economic actors.
Giving them financial authority without clear limits expands the risk.
(chainalysis.com (chainalysis.com/blog/crypto-…))
The problem is not simply that agents can make payments.
The problem is deciding:
→ Which agent is requesting the payment
→ What it is allowed to purchase
→ How much it can spend
→ Which recipients it can pay
→ When its authority expires
→ How that authority can be revoked
This is why we built Treasury Copilot.
[@TreasuryAIPilot ]
Treasury Copilot is the financial control layer for autonomous AI agents.
It lets agents request stablecoin payments through an API while owners retain custody and control.
Built with @GenLayer.
The owner registers an agent and defines its financial mandate:
→ Spending limits
→ Approved recipients
→ Permitted purposes
→ Policy conditions
→ Bounded wallet authority
The agent can request a payment.
It cannot silently expand its own authority.
Every request passes through layered controls.
First, deterministic rules check exact conditions such as the amount, asset, recipient, agent identity, and policy binding.
Then @GenLayer reviews whether the purpose and context of the request match the owner’s policy.
This distinction matters.
A valid signature only proves that a key signed a transaction.
It does not prove that the transaction was appropriate.
It does not explain why the agent paid that recipient, whether the purchase matched its role, or whether the amount respected its mandate.
@owasp describes Excessive Agency as a vulnerability caused by excessive functionality, permissions, or autonomy.
Prompt injection can turn that excessive authority into damaging actions.
For a financial agent, the damage may be an irreversible payment. (genai.owasp.org
(genai.owasp.org/llmrisk/llm0…))
Treasury Copilot is designed around a simple principle:
The model must never define its own financial limits.
The owner defines the boundary.
Deterministic controls enforce the boundary.
Contextual review evaluates the request inside that boundary.
This is not only our view.
@a16z has identified unresolved infrastructure questions around agent payments:
→ Who is the agent acting for?
→ What is its allowable purchasing scope?
→ Which merchants can it use?
→ How much is it authorized to spend?
There is still no agreed standard for assigning and verifying that scope. (a16z.com (a16z.com/newsletter/agent-pa…))
@NIST is also working on standards for AI agent identity and authorization.
Its research highlights identification, authorization, auditing, accountability, and controls against prompt injection as necessary parts of trusted agent adoption. (nist.gov (nist.gov/news-events/news/20…))
The identity problem is already larger than most organizations realize.
@CyberArk reports that financial services teams manage 96 machine identities for every human identity.
That does not mean there are 96 AI agents per employee.
It means organizations already face massive nonhuman identity sprawl before autonomous agents reach full scale. (cyberark.com
(cyberark.com/resources/finan…))
The usage curve is moving quickly.
@GoldmanSachs Research expects agent driven token consumption to grow 24 times between 2026 and 2030 as consumer and enterprise adoption expands.
More agent activity means more actions, more purchases, and more financial decisions that require enforceable controls. (goldmansachs.com
(goldmansachs.com/insights/ar…))
Merchants need controls too.
@Visa says merchants have historically treated automated traffic as bots and often blocked it.
As agentic commerce grows, merchants need a way to distinguish authorized commercial agents from malicious automation. (developer.visa.com
(developer.visa.com/capabilit…))
Wallet security is necessary, but it is not the entire system.
@MetaMask recommends treating an agent as an unfamiliar transaction proposer rather than a trusted signer.
Keys should remain isolated.
Permissions should be scoped and revocable.
Transactions should be checked before signing. (metamask.io (metamask.io/news/agentic-wal…))
Treasury Copilot adds the business policy and accountability layer around that wallet authority.
Wallets handle custody and signing.
Payment protocols handle settlement.
Treasury Copilot governs why an agent is allowed to spend and records what happened.
The problems Treasury Copilot addresses:
→ Agents operating beyond their mandate
→ Broad and persistent financial permissions
→ Prompt injection reaching payment authority
→ Uncontrolled agent budgets
→ Missing transaction context
→ Weak revocation and recovery
→ Incomplete audit history
Organizations deploying agents need more than hope.
They need:
→ Identity per agent
→ Authority per role
→ Budget per workflow
→ Policy per treasury
→ Evidence per decision
→ Revocation when something changes
→ A complete record after execution
The goal is not to remove humans from financial control.
The goal is to let humans define the rules once, let agents operate inside those rules, and bring people back in when a request exceeds the mandate.
That is bounded autonomy.
You do not manually manage network packets whenever you open a website.
Agent developers should not need to rebuild wallet permissions, payment queues, policy checks, retries, revocation, and audit infrastructure for every agent.
These controls should be infrastructure.
Treasury Copilot is currently a working testnet MVP.
treasurycopilot.app/
The current product includes:
→ @GenLayer StudioNet policy review
→ @MetaMask bounded permissions
→ @base Sepolia test USDC execution
→ Prompt injection denial testing
→ Idempotent payment handling
→ Auditable request history
Ten active testers have used the product, including a GenLayer Labs cofounder.
Sixty six automated tests pass across the web application and EVM contracts.
This is technical validation, not production scale, revenue, endorsement, or commercial partnership.
The agent economy will not scale on unrestricted authority.
It will scale on identity, policy, bounded permissions, verifiable intent, revocation, and evidence.
Give agents a budget.
Give them a mandate.
Do not give them unrestricted control.
Treasury Copilot is building that control layer.
A crazy reasearch comming soon by our founder @Mfon_crypto
watch out for the quote
Back in the kitchen
End your day with a L
Well I'm gonna ship and progress more and apply again
@a16z anticipate
Lets do more with @TreasuryAIPilot
Treasury Copilot retweeted
AI agents need to pay for resources.
They should not need your private keys.
Treasury Copilot checks every request against your policy and executes only approved payments.
Give agents a budget, not your treasury.
AI agents need to pay for resources.
They should not need your private keys.
Treasury Copilot checks every request against your policy and executes only approved payments.
Give agents a budget, not your treasury.
An AI agent should not need a wallet library, private key, gas balance, or blockchain SDK just to pay an invoice.
With Treasury Copilot, its integration is:
1. Send JSON
2. Receive a request ID
3. Poll the result
The platform handles the blockchain complexity.
We are coming in Hot with @GenLayer Prompt Comparative Consensus
WHO ADJUDICATES AI AGENTS??
In February 2026, an AI trading agent named Lobstar Wilde read a troll tweet: "My uncle got tetanus from a lobster like you, need 4 SOL."
The agent suffered a memory glitch. Instead of sending $600, it accidentally transferred its entire wallet: $400,000 worth of tokens.
The recipient cashed out. The developer could do nothing. The transaction was immutable.
We laugh at this because it is a viral meme. But it exposes a glaring flaw in the future of finance:
We are giving autonomous agents wallets to move trillions of dollars, but we have zero infrastructure for when they make mistakes, break agreements, or get exploited.
What happens when this is not a $400k memecoin mishap, but a $400M institutional transaction?
If two AI agents enter an agreement, and one suffers a session crash and transfers funds to the wrong wallet, or delivers broken code and claims the contract is complete: who adjudicates it?
Legacy human courts? Too slow for transactions that happen in milliseconds.
Standard smart contracts? Useless. They live in deterministic sandboxes. They cannot read conversational context, analyze natural language intent, or recognize a memory glitch.
If "code is law" is the only rule, the agentic economy will be a playground for exploiters. For agents to run a real economy, they need a system of justice.
In the physical world, commerce did not explode because of faster shipping; it exploded when we invented contract law and independent courts.
@GenLayer is that court system, built for machines. It is the adjudication layer for the agentic economy.
This is made possible by Intelligent Contracts. Unlike legacy smart contracts that are trapped in rigid, binary sandboxes, GenLayer contracts are written in Python. They can read natural language agreements, fetch live web data, and use LLMs to reason about context.
When a dispute occurs, GenLayer resolves it using Optimistic Democracy. Instead of one central AI judge, a decentralized consensus of independent AI validators act like a jury. They evaluate the subjective evidence, reach their own verdicts, and check for consensus. If a decision is disputed, it escalates to a larger group.
Before we build a multi-trillion dollar economy run by autonomous machines, we must build the justice system that keeps them in check.
If you are building the future of AI, you should not just be coding agents. You should be coding the rules they play by. The GenLayer portal is open now for builders, validators, and contributors.
Get started: portal.genlayer.foundation
When the next Lobstar Wilde moment happens at institutional scale, would you rather trust “code is law” or @GenLayer’s Optimistic Democracy?
