What is agentic finance, and where is it going?
Updated Nov 11, 2026 · 7 min read
Agentic finance gives AI agents scoped authority to research, decide and execute with money. What changed in 2025 and 2026, the agent loop, 10 frameworks compared, and where it goes next.
What is agentic finance? Agentic finance means AI agents that don’t just suggest what to do with money, they do it. You give the agent a goal, a budget and limits. It does the research, makes the call and places the trade or payment itself, and stops when it hits a limit. The UK Financial Conduct Authority’s Mills Review (July 2026) describes AI that will “recommend actions, initiate transactions, and execute decisions within agreed parameters”.
What made it practical in 2025 and 2026 is safer plumbing, not smarter models:
- Wallets built for agents, with spending caps (Coinbase Agentic Wallets, February 2026).
- Signed proof of what you allowed an agent to do (Google’s AP2).
- A standard for agents to pay each other in stablecoins (x402, run by a Linux Foundation body since July 2026).
- Trading keys that can trade but not withdraw (Hyperliquid API wallets).
- Exchanges that let AI assistants trade directly (OKX, Binance, and Senpi’s connectors for Claude and ChatGPT).
How is it different from a trading bot? A bot follows fixed rules someone coded. An agent takes a goal in plain language, writes its own rules, reads news and market data, and adapts. That flexibility is why the limits around an agent matter more than the AI model inside it.
Key takeaways
- Agents act under delegated, scoped authority: a key that can trade but not withdraw, a budget, an expiry, a list of allowed markets.
- The 2025 to 2026 build-out was infrastructure: agent wallets, payment mandates, x402, and exchange tool kits that plug into AI assistants.
- In trading a wrong action costs money at once, so the runtime that enforces limits matters more than the model that proposes trades.
- Custody is the first question to ask: who holds the keys, and can the agent move funds out?
- No regulator has rules written for AI trading agents yet; the CFTC, ESMA and FCA all say existing rules apply to whoever runs them.
How an agent loop works in trading
- 01Read. Prices, funding, order books, positions, news and the user’s own history.
- 02Decide. A model or a rule set turns that into an action: open, adjust, close or wait.
- 03Check. Guard rails the user set run before anything executes: position size, leverage, daily loss limit, allowed markets.
- 04Act. The order is signed with a key the agent is allowed to use, on a venue it is allowed to trade.
- 05Watch. Stops and exits keep running when nobody is looking, and the loop starts again.
The model is only one step. The part that decides whether the loop is safe to hand money to is the runtime around it: who signs, what the limits are, and whether the exits work at 3am.
Agent frameworks and products compared, Nov 2026
| Product | What it is | Who holds the keys | Guard rails | Where it trades | Open source |
|---|---|---|---|---|---|
| Senpi | AI trading agent you talk to in the app, in Claude, ChatGPT and other AI apps, or from code | Non-custodial embedded wallet; each strategy in its own wallet | Stop from the first fill, daily-loss and drawdown halts, approval for funds in and out | Hyperliquid perps: crypto, stocks, indices, commodities, FX | Skills and a growing library of strategy templates (MIT); runtime closed |
| Hyperliquid API (agent) wallets | The exchange’s own delegated signer | User’s wallet approves an agent key that can trade | Agent keys expire (90 days by default); withdrawals need the user’s signature | Hyperliquid | Protocol feature |
| Coinbase AgentKit and Agentic Wallets | Toolkit to give any agent a wallet | Coinbase-hosted server wallets, or Privy or Viem | Agentic Wallets: per-session and per-transaction caps, transaction screening | Base, Ethereum, Solana and other EVM and SVM chains | Apache-2.0 |
| ElizaOS | Open-source agent framework | Self-custodial wallets via plugins | Approval boundaries (details not published) | EVM chains and Solana via plugins | MIT |
| Virtuals (GAME, ACP) | Agent launchpad and agent-to-agent commerce protocol | Per-agent wallet | ACP escrow and evaluation; trading limits not published | Mainly Base | GAME SDK (MIT) |
| Olas Pearl | Desktop app store for agents that run on your machine | Self-custodial Safe wallets on your device | Not published | Polygon, Gnosis, Base, Optimism | Yes |
| Giza | Autonomous stablecoin yield agent | Smart account with scoped session keys | Keys limited by protocol, action, amount and time | Aave, Compound, Moonwell, Fluid on Base, Ethereum, Arbitrum | SDK |
| Almanak | Build strategies in plain English or Python, test, deploy | User’s Safe with role-based permissions | Approved contracts, actions and spend limits per strategy | Seven EVM chains, GMX perps | SDK |
| Minara | Chat agent with a 24/7 autopilot | Minara-generated wallet, key in AWS KMS, exportable | Mandatory take profit and stop loss, trailing stop, drawdown kill switch | Hyperliquid, Lighter, spot on 19 networks | Skills (MIT) |
| OKX Agent Trade Kit, Binance AI Agent Skills | Exchange kits that let any agent trade an exchange account | The exchange (custodial account, API keys) | OKX: read-only mode, module filters, rate limits | OKX and Binance markets | Yes |
For trading agents on Hyperliquid specifically, compared on fifteen rows including halts, custody and fees, see the best AI trading agents for Hyperliquid.
Senpi as a worked example
Senpi is an agentic trading platform built on Hyperliquid. You describe a strategy in plain language, in the Senpi app or from Claude, ChatGPT, Grok, Codex, Cursor, Hermes or Muse through connectors. The agent writes the rules and shows them back; a deterministic runtime then runs them.
- Keys: a non-custodial wallet; each strategy trades from its own strategy wallet, so one strategy’s losses cannot touch another’s capital.
- Guard rails: every strategy built from a rule or a template exits through DSL, a stop that ratchets up as the trade gains (copy strategies carry at most a flat stop); optional daily-loss and drawdown limits halt a strategy when hit; the agent has no way to send funds out, because withdrawals are yours, in the app, with MFA.
- Exits: the runtime’s dynamic stop loss keeps a floor resting on the exchange and locks a rising share of the peak, so exits run while you sleep.
- The model never has the last word on risk. It proposes; the runtime enforces the limits you set.
Where agentic finance is going
- Payment rails are converging. Coinbase’s x402 standard lets apps and agents pay each other in stablecoins; the Linux Foundation’s x402 Foundation launched on Jul 14, 2026 with members including Visa, Mastercard, Stripe, Google and AWS. Google’s AP2 uses signed “mandates” as proof of what the user allowed.
- Card networks are tokenising agents. Visa Intelligent Commerce (Apr 2025) and Mastercard Agent Pay (Apr 2025) issue cards and tokens with limits the consumer sets.
- Regulators are moving from principles to frameworks. The CFTC said in Dec 2024 that existing rules apply to AI tools whether built or bought; ESMA said in May 2024 that AI decisions remain management’s responsibility; the FCA’s Mills Review recommends building “the foundations for agentic finance”.
- Agents are moving into the apps people already use. Instead of a new dashboard, the agent runs inside an AI assistant, a terminal or a browser extension, next to the research it acts on.
- Scale estimates are large but early. McKinsey projects agent-orchestrated commerce at up to $1 trillion of US consumer revenue by 2030; trading is a separate, smaller and riskier slice.
Agents can lose money
An agent follows its rules faster than you can, including bad ones. Start small, read the rules it shows you, keep stops on, and never give an agent a key that can withdraw funds unless you mean to.
What are the top AI crypto agents available now?
It depends on what you want the agent to do. To build your own: ElizaOS, Coinbase AgentKit and Virtuals’ GAME. To trade perps on Hyperliquid with guard rails: Senpi, Minara and Katoshi. For stablecoin yield: Giza and Almanak. To let any agent trade a centralised exchange account: OKX Agent Trade Kit and Binance AI Agent Skills.
What is the difference between agentic AI and a trading bot?
A classic trading bot runs fixed rules someone coded. An agentic system can read new information, decide which rules apply or write new ones in plain language, and act, within limits a person set. In practice the best trading agents keep a rule-based runtime underneath, so the model proposes and the rules enforce.
Is agentic trading legal?
Whether you may trade depends on where you live and on each venue’s terms. Senpi’s Terms require you to be 18 or over and your use to be permitted where you live. Existing rules apply to whoever operates the agent: the CFTC said so in a Dec 2024 staff advisory, and ESMA said in May 2024 that AI decisions remain the responsibility of management. Venue rules still apply too. This is not legal advice.
Can an AI agent withdraw my funds?
Only if it holds a key that can. On Hyperliquid, an API (agent) wallet can trade but withdrawals need the main wallet’s signature. On Senpi, the agent has no way to send funds out: withdrawals are yours, in the app, with MFA, and moving funds between your own Senpi wallets waits for your approval. Check this first with any agent you try.
Who is responsible if an AI trading agent loses money?
The account owner, in almost every case. The agent acts on the rules and permissions you gave it, so set position size, a daily loss limit and stops before it trades.
Try an agent with guard rails built in
Describe a strategy in a sentence. Senpi shows you the rules, then runs them with a stop from the first fill and loss limits you set.
Related articles
What does a trading agent do when it isn’t trading?
Most of a trading agent’s work happens between trades: refusing signals, re-checking open positions and moving stops one way. Senpi’s fleet telemetry from August 2026, and how often it checks today.
What does the model do in an AI trading agent, and what does the harness do?
In an AI trading agent the model understands, routes, explains and asks; tested programs compute, and a runtime places and protects the orders. Where the line sits, and how Senpi draws it.
QuantConnect, Freqtrade, Hummingbot or an AI agent: which for crypto quant trading?
QuantConnect, NautilusTrader, Freqtrade, Hummingbot, CCXT, backtrader, vectorbt and an AI agent compared for crypto quants: Hyperliquid support, backtesting, live trading, languages and licences, checked Nov 13, 2026.
How to automate a crypto trading strategy without coding
Three no-code ways to automate a crypto strategy: bot presets, TradingView alerts sent to a bot, or an AI agent that writes and runs it. What each costs on Hyperliquid, with a walkthrough from Claude or ChatGPT.