Back From Lisbon
There was no developer meeting last week. We were all in Lisbon for Hack Meridian and the Meridian conference. If you missed the talks and workshops from Hack Meridian (Vlad, Elliot, Nicole, me and others), they are on the SDF YouTube streams tab. Some numbers from Hack Meridian 2026: about 330 to 350+ builders checked in, from about 40 countries, and they submitted 165 projects. My guests this week were two of those builders.What Koul Is
Koul is a team from Turkey and one of our latest SCF Instawards recipients. Koul is automated trading infrastructure for AI agents on Stellar. Its README calls it a conditional execution engine. You write a rule once. For example: “if my loan health drops below 1.2, repay the debt”, or “if BTC drops below 110k, move half my position to USDC”. The rule lives on chain as Soroban contract state. It executes from your own wallet through a scoped key. Your funds never leave your wallet, and you can revoke access with your passkey at any time. Koul is live on Stellar testnet. It is built with Soroban, passkeys and OpenZeppelin smart accounts, and it integrates with XOXNO (live on mainnet) and the Reflector oracle.Meet the Founders
My guests were the co-founders, Ahmed Murshed and Atahan Yıldırım. Ahmed is a final-year computer engineering student in Turkey, with about four years in blockchain and many hackathons behind him. He built Sembol, a library that creates OpenZeppelin smart accounts with passkeys (Face ID, fingerprint) from one import. Ahmed said Sembol is also an Instawards winner. Atahan worked on a conditional execution layer on Stellar: “you have rules, and you allow it to run”. They met through their university blockchain club. Both then worked for about a year at a US-based startup that builds trading infrastructure, where they learned what traders want. The Koul idea came up at a hackathon in Istanbul, two to three months before our call. Sembol plus the execution layer gave them on-chain strategies without giving custody to bots or agents. Before Hack Meridian, Koul did only lending and borrowing. At Hack Meridian they added swaps. Atahan called Hack Meridian and Meridian the best event he has attended.A Testnet Trading Competition
After the hackathon, Koul launched a trading competition on testnet. It is free to enter, with a prize pool of $300 in USDC. You describe a strategy, either manually or in plain language to the built-in AI, and Koul turns it into a strategy that runs on chain. On the call they had over 110 participants. The competition launched about six days before the call and ran until November 6. It is testnet, so nobody risks mainnet funds. And, as I said on the stream: not financial advice.Live Demo
Atahan walked through the app:- Create a wallet. You do not need an existing wallet. A fresh wallet gets 100 USDC and about 10K XLM on testnet.
- Compete screen. An overview and the leaderboard.
- Autopilot page. This is the main page. Pick a template strategy, build one manually, or ask the AI. The demo prompt was “buy the XLM dip”. The AI returns a rule, and you can change it, for example from 10 USDC to 33 USDC.
- Conditions and actions. Conditions can use token price, the tokens you hold, and pool APY. Actions today are swaps and liquidity pools. Swaps go through the XOXNO swap aggregator, which finds the best route.
- Passkey confirmations. Starting an autopilot asks for a few passkey confirmations, once. After that, no confirmation is needed when a rule fires. Edit, pause, stop and delete need the passkey.
- Custody. Usual agent setups give the agent your keys or a funded hot wallet. If the agent is open to the internet, anyone can manipulate it and drain the funds. Off-chain bots still need key management.
- Latency. An agent reads an oracle or the chain, evaluates, sends a transaction, and waits for settlement. By then the price advantage is gone.
Questions During the Demo
- Backtesting? On the roadmap, probably after perp support.
- Sentiment data from third parties? Possible, since any data source can plug in. Rules are static today (wallet balance, token price). They plan to add metrics and technical formulas, like the TradingView scripts that traders use.
- Create liquidity pools or find arbitrage? Not today. It is on the roadmap.
- Which smart accounts? Sembol, which uses the audited OpenZeppelin smart accounts.
Architecture
On the morning of the call there were over 100 autopilots. The flow: you write a strategy in plain language, Koul compiles it into rules, and a router contract stores the autopilot. The components:- Web app at koul.me.
- Keeper. The only off-chain part. Contracts do not run by themselves, so the keeper calls
tickon an interval. It simulates first and submits only when the contract says an action applies. It handles liveness only. Two keepers run today, and the plan is to make the keeper permissionless, so anyone can run one. - User smart account. One rule holds your passkey. Another holds Koul’s agent key, bound to a policy contract, with an expiry. Koul cannot withdraw your funds. It can only run your strategies.
- Router contract. It stores the rules, evaluates them in order, and runs at most one action per tick: the first rule that is true. Decide and act happen in the same transaction, so the step is atomic and there is no latency between the decision and the action. The router reads Reflector prices and calls XOXNO lending and the swap aggregator.
- The keeper simulates. If nothing applies, it stops.
- For a swap, the keeper gets a route.
- The keeper submits the tick, signed with the agent key.
- The policy checks the root call and every sub-call. Anything off the allowlist fails.
- The router re-reads state, runs the first true rule, and emits an event. The app reads the event.
- Perp DEX integrations, for example “open a 12x BTC long, cut to 5x and add collateral if X happens”.
- More oracles, to cut latency.
- Cross-chain support, which they called the hardest item.
- An open rule format, so other AI agents can use Koul as infrastructure only.
Latency
Latency today is about 10 to 15 seconds from condition to execution. Their plan to bring it down:- About 7.5 seconds is on the keeper side. A fix ships the day after the call, to under 1 second.
- The RPC adds 2 to 3 seconds. The plan is a paid high-performance RPC, or a keeper on a VPS next to the RPC. Together with the keeper fix, that takes the combined 8.5 seconds down to 2 to 3 seconds.
- Fewer round trips per tick, and more parallelism.
- Pull oracles instead of push oracles for price rules. The current oracle updates every five minutes. A pull oracle can bring this to under a second.
- Ledger close, about 5 seconds, is the floor. The target is about 6 to 7 seconds end to end.
My Feedback and Q&A
My view: running the keeper on GitHub Actions explains much of the 7.5 seconds. I suggested Cloudflare dynamic workers for the off-chain parts (I measured about 500 ms in my own demos), plus a low-latency feed such as Pyth Pro. I think about 6 seconds in total is possible.- Where are the keys? The passkey stays in the device’s secure enclave and never leaves it. Koul only holds a scoped, expiring execution policy that you can review.
- How hard was it to build on XOXNO? Easy, they said. The XOXNO repo has a skills folder for each part of the protocol, so they pointed their AI at the skills and the contracts. XOXNO skills are also in the Stellar dev skills.
- When mainnet? First, lower latency and better scalability. Then features such as perps. They also want an audit: they do not see the contracts as security-critical, but they still want one. The target is mainnet “in a couple of months at most”. They will rank features by the feedback from the competition and from this stream.