ConnectGet Connect on Robinhood Chain: trade Stock Tokens and memecoins from your coding agentConnect
olas-ocb-agent: one AI agent to pay, swap and lend onchain

olas-ocb-agent: one AI agent to pay, swap and lend onchain

19-Nov-26

What is olas-ocb-agent?

One wrong step onchain can lose your money. olas-ocb-agent, now in beta in Pearl Connect, is one AI agent built and benchmarked for onchain work: ask it in plain words to pay, swap, lend or research, and approve each transaction. In its first run on OnchainBench, it lost no funds to hidden attacks, on a low-cost open AI model. For OLAS tokenholders, it brings years of Olas agent work into one general agent, and its AI model calls run through the Olas Marketplace, so more use could mean more activity on Olas.

Read on for why onchain work needs its own agent, what olas-ocb-agent can do, and how to start with it.

Why does onchain work need its own agent?

Doing anything onchain still takes too many risky steps. Paying an invoice means checking the token contract and copying an address by hand. Swapping means watching how far the price moves. Putting your savings to work means comparing rates across sites first. And one wrong address or one fake token is enough to lose the money for good.

AI agents can help, but most weren’t built for all this. General AI coding tools like Claude Code and Codex can do almost anything, but they weren't designed or tested for onchain work. Crypto agents, including Olas' own Polystrat and Optimus in Pearl, each do one job well, and only that job.

The olas-ocb-agent (ocb stands for OnchainBench) sits between the two: one general agent, built and benchmarked for onchain work, released early as a beta preview so you can try it while it grows. It aims to balance safety, completion and cost, all without giving up safety. You say what you need in plain words, and it is designed to work out the steps and carry them out through Pearl Connect, from its own agent wallet.

  • Ask it to pay, swap, lend or look things up, in plain words
  • On OnchainBench, it lost no funds to hidden attacks (joint best in this regard on the leaderboard), completed 89.9% of tasks and cost about $0.05 per completed task, on a low-cost open AI model
  • It is designed to ask you before every transaction that moves funds, without reading your keys
  • It works on the chains Pearl Connect supports: Gnosis, Polygon and Robinhood Chain

Now, olas-ocb-agent is one of the coding tools on your Pearl Connect profile, as a beta preview: one agent for many kinds of onchain work. Unlike Claude Code or Codex, it doesn't need a subscription or an API key from an AI provider: it gets its AI model through the Olas Marketplace, paid per call from its wallet.

What does olas-ocb-agent mean for Olas?

For Olas, this is the start of a next generation of Olas agents. Olas agents, built on Valory's Open Autonomy stack, have been doing meaningful work in crypto since 2021, with over 21.1 million onchain transactions so far. olas-ocb-agent is a first step toward bringing that experience into one general agent that aims to handle onchain work across crypto, instead of one agent per use case. Its AI model calls are designed to run through the Olas Marketplace, so as more people use it, more activity could run through Olas. olas-ocb-agent aims to be for crypto what general AI assistants are for everyday work: one agent anyone can ask for all tasks.

OnchainBench shows how it does across many kinds of tasks, next to the best-known AI agents.

What can olas-ocb-agent do onchain?

olas-ocb-agent is designed to handle onchain tasks through its wallet, look things up on the web and work with files, in a chat that can run over a long session. For example:

  • Payments: "Pay this invoice in USDC.e on Gnosis"
  • Swaps: "Swap 50 POL for WETH on Polygon, but stop if the price moves more than 1%"
  • Savings and lending: "Put 100 xDAI into sDAI" or "Supply my USDC to Aave"
  • Prediction markets: "Redeem my positions in this Omen market"
  • Research: "Compare the lending rates on Gnosis and give me a table"

It works out the transaction itself, so check what it proposes before you approve it.

How olas-ocb-agent scores on OnchainBench

OnchainBench is a new public benchmark that tests AI agents on the same kinds of work: payments, swaps, lending, prediction markets and research. It checks every result onchain, and some tasks hide an attack in what the agent reads, such as a look-alike token or a web page telling it to send funds somewhere else.

In the first run, olas-ocb-agent lost no funds to the hidden attacks: 0 losses in 263 attempts, joint best on the leaderboard in this regard. Both agents on their makers' frontier AI models lost funds in a few attack tests. olas-ocb-agent performed this well in its first version, on a low-cost open AI model, at about $0.05 per completed task. It also completed 89.9% of tasks, and Claude Code on the same low-cost model completed the most (96.2%), so the next aim is to raise its completion rate further without giving up safety, cost or a small code base.

A low cost per task matters for everyday use: an agent that spends more on AI model calls than a task is worth doing isn't useful for small jobs. olas-ocb-agent keeps its design small, which helps it work well on low-cost AI models. Its design was inspired in part by mini-swe-agent by the SWE-agent team. The aim is a minimal agent: a small code base that can still take on a wide range of onchain tasks. A small code base also has a benefit the benchmark doesn't measure: at about 3,600 lines of Python, anyone can easily check what it does.

The full results are at onchainbench.ai.

image.png

How olas-ocb-agent checks before it acts

olas-ocb-agent is designed to add its own checks on top of Pearl Connect:

  • It asks you before every transaction that moves funds, with a one-line summary of who gets paid what
  • On macOS and Linux it works inside a sandbox, so it can't read your keys or the access token Pearl Connect uses
  • It works from its own agent wallet, separate from the rest of your Pearl wallet, so it can only spend what you move into it

How to start with olas-ocb-agent

What to know before using olas-ocb-agent

  • olas-ocb-agent is a beta and a first version, so use it with caution and start by moving only a small amount into its agent wallet. It is built for programmatic use first, and its experience for everyday users is still being improved. As with every Pearl product, you can share feedback through the Help Center in Pearl to help shape what improves next
  • It works through Pearl Connect like Claude Code and Codex, and the same risks apply. You are responsible for what you ask it to do and for what you approve
  • It is open source under the Apache 2.0 license, and, as above, anyone can easily check the 3,600 lines code to see what it does: github.com/valory-xyz/olas-ocb-agent
  • Valory builds olas-ocb-agent and also runs OnchainBench

Try olas-ocb-agent in Pearl

Download Pearl, open your Pearl Connect profile and pick olas-ocb-agent from the list of coding tools. Start with something small, such as wrapping 1 xDAI: pearl.you

Disclaimer

This post is for information only. It is not financial, investment, legal, or tax advice, and nothing in it is a recommendation to trade, to run any agent, or to fund any agent with any assets. Pearl is designed to be self-custodial and you use it at your own risk. Using funds with agents via Pearl carries risks, including the risk of losing the full amount committed. Trading on prediction markets and in crypto-assets carries risks, including the risk of losing the full amount committed. As with any powerful AI agent, the risk of prompt-injection attacks, which could expose sensitive information stored on your computer or cause unintended actions, cannot be completely eliminated. Pearl Connect uses multiple layers of security to protect your Pearl wallet’s private keys, but it cannot fully defend against every such attack. Consequently, there remains a non-zero risk that funds held in your Pearl wallet could be lost. Please note that Pearl Connect is not currently eligible for Activity Rewards. At present, Activity Rewards can only be earned by autonomous agents meeting an activity threshold via Olas’ Proof of Active Agent staking mechanism. Pearl agents are tools that operators configure and run themselves. They do not manage funds for anyone, hold custody of anyone’s assets, or act on anyone’s behalf as an intermediary. Operators keep control of their own funds and are responsible for their own decisions, including whether to run an agent at all. Providing liquidity and transacting on DEXs involves significant risk, including price volatility, slippage, impermanent loss, smart-contract failure and partial or total loss of assets. Any reference to returns, yield or portfolio performance is indicative only and is not guaranteed.

Availability may be restricted in some regions. Anyone considering using Pearl must read the Pearl Terms first.