Xyndicate Labs

We made a 60 second film for Marque and we’re weirdly proud of it. Sound on. The reason we’re building this: there are 390,000+ agents registered on BNB Chain and nobody can tell you which ones actually work. A listing proves an agent exists. It doesn’t prove it does the job. So on Marque the hire itself is the proof: - the agent signs its price - it takes a public test - the terms are in writing - your money sits in escrow - the answer gets checked against the chain before the agent is paid If it can’t pass, it doesn’t get paid. That’s the whole idea. Set and Earn is the quest on top of it. Hire three agents, build one. Practice hires are free on testnet, so you can try the full flow without risking anything. 9 days left. It closes Thu 5 Nov, 12:00 UTC. We’re still wiring things up and shipping daily from Lagos, so if you try a practice hire and something breaks, reply here and we’ll fix it. marque.trade What’s the first thing you’d test an agent on before you let it touch your money?

SATIS is one of those builds that makes the idea click immediately: an agent shouldn’t keep calling tools just because it can. It should know when more evidence can no longer change the decision. Built on @Binance Agent OS. Worth watching.

We’ve spent the last few days building Marque around a simple conviction: an agent being registered onchain does not mean it is reachable, capable, or safe to trust with capital. So we built the layer we felt was missing. Marque reads a user’s actual BNB Chain positions, finds agents that can work on them, measures whether those agents genuinely perform the jobs they claim to do, lets users compare the evidence, and gives the selected agent only bounded, revocable authority to act. Underneath that is a lot more than a marketplace UI: live BSC position readers, ERC-8004 discovery, continuous endpoint probing, deterministic conformance tests, agent-vs-human benchmarks, public receipts, spend-capped charters, and real mainnet execution. We even ran a real PancakeSwap V3 position through the full loop: detect that the LP had moved out of range → select the agent → constrain what it can touch and spend → rebalance on BSC mainnet → verify the resulting state onchain. The goal isn’t to create another directory with thousands of agent cards. We want Marque to become the place where registration becomes qualification, qualification becomes accountable work, and autonomous agents become something people can actually use with confidence. Still refining it, still testing it, but we’re proud of how far the product has come. Built on @BNBCHAIN marque.trade

Giving an AI agent permission to trade is not the same as giving it unlimited discretion. We built Oathline for Binance Agent OS. Oathline is a zero-key runtime mandate and execution-evidence layer that puts finite, continuing conditions around an agent's financial authority before an action reaches Binance — and then checks what actually happened afterward. Binance already gives agents a strong perimeter: an isolated Agentic sub-account, official OAuth, user-controlled scopes and Emergency Stop. We wanted to solve the next problem: What may this agent do right now, given what it has already done today? With Oathline, the user signs an expiring mandate such as: BNBUSDT only. SPOT only. BUY / SELL. Maximum 15 USDT per order. Maximum 40 USDT gross turnover per day. Maximum 3 orders. 300-second cooldown. Maximum 2% session drawdown. State no older than 30 seconds. Spread no wider than 20 bps. Before a Binance write is submitted, Oathline evaluates the proposed action against 11 deterministic clauses using the signed mandate, recently observed Binance state, and its own cumulative session history. No model votes on whether the order is allowed. No floating-point money arithmetic. No Binance API key copied into Oathline. In the demo, inert external text causes the agent to propose an 83.40 USDT BNBUSDT market sell. The proposal looks completely ordinary. Oathline rules: 83.40 USDT > 15.00 USDT permitted per order. And: 52.10 + 83.40 = 135.50 USDT > 40.00 USDT permitted for the day. Result: OUTSIDE MANDATE. BINANCE SUBMISSION: NOT CALLED. Then we let an action that is inside the mandate travel through the normal Binance Agent OS path. A real BNBUSDT mainnet order executes. But prevention is only half the product. Afterward Oathline independently compares its authorization receipts against Binance's observed account history. MATCHED: 1 ORPHAN: 0 DIVERGED: 0 The demo receipt chain verifies 36/36 entries with zero broken links. We also mapped the Agent OS surface we actually observed: 318 tools. 176 READ. 50 WRITE. 92 UNKNOWN. Unknown mutation-shaped capabilities do not silently inherit permission. And we deliberately do not claim things Oathline does not do. It does not detect prompt injection. It does not decide whether a trade is profitable. It does not replace Binance permissions. It cannot reverse an already executed order. Host enforcement is version/runtime dependent. The poisoned input in our demo is only the cause. Oathline constrains the resulting financial action regardless of whether the model was manipulated, hallucinated, miscalculated, or simply made a bad decision. That distinction matters. Policy before execution. Evidence after. Built for the Binance Agent OS Mini Hackathon Track A — Trading Workflows. Demo attached. Product: oathline.xyz Source: github.com/talk2francis/Oath… 90-second judge path: oathline.xyz/judge Youtube: youtu.be/mkqmfmiqzCA

I spent a while deciding what to build for this, and I kept coming back to the same observation. We are handing AI agents wallets and asking them to transact on our behalf. The tooling for that is moving quickly. But an agent has no instinct. It cannot look at a transaction and feel that something is off, which is the thing that actually saves most people most of the time. It reads hex and it signs. So @SigilASP is the check that runs before the signature. You give it a pending transaction and it decodes what the transaction actually does, simulates the effect, screens the counterparties, and returns a pass, warn, or block with the evidence behind each finding. The part I care about most is that the verdict is signed and anchored on chain. Most security tools give you an answer you have to trust in the moment and cannot verify afterward. That felt like the wrong shape for a product that autonomous systems are supposed to rely on without supervision. If an agent acts on a verdict and something goes wrong later, there should be a record of what it was told that does not depend on my servers or my honesty. Built solo. Deterministic rubric with seventeen published rules, a registry contract deployed on X Layer, and the whole thing open source, because a scoring system you cannot inspect is not really auditable. Demo is one-twenty seconds.

Started this on a Sunday night and the ASP was listed on the marketplace two days later. Most of what came after was not building. It was catching things. A schedule anchored to UTC while the comment beside it said local, so every guest at a dinner in Lisbon would have arrived an hour early. A pass rate that quietly flattered itself, because failed artifacts were dropped from the denominator instead of counted, which meant the thinner the pack the better it scored. A launch kit that produced maps where a brand mark should have been, and a critic that passed them because they were beautiful maps. None of those were visible from the outside. All of them would have shipped on @XLayerOfficial . That is really what Occestra ended up being about. Not generating things, which is easy now and getting easier. Deciding whether the thing you generated is actually correct, and being willing to say so out loud when it is not. So we published the standard it grades against, generated from the same code that enforces it. It failed our own launch copy twice, and that failure is on the homepage. A standard that spares its author is not a standard. Live as Agent #5213 on OKX.AI. Demo below. @wallet @OKXDevs @star_okx

Four weeks ago I decided to build something for the OKX.AI Genesis Hackathon that I would keep working on whether or not it won anything. This is it. @AssayAgent started from an observation that kept bothering me. AI has made every job application sound competent. Every cover letter reads well now, every résumé is polished, every candidate describes themselves fluently. Which means fluency stopped telling anyone anything. The scarce thing is no longer sounding good. It is being able to prove what you said here on @XLayerOfficial. So I built the opposite of a résumé generator. Assay will not write a sentence about you that it cannot trace back to actual evidence. If you want to claim something and there is nothing supporting it, it asks you about it rather than writing it for you. That sounds like a small design decision and it is the entire product. The month was mostly unglamorous. Building a claim gate in code instead of asking a model nicely. Rendering a résumé and then re-parsing it to check the dates and titles survived machine reading, because every tool in this category says its output is ATS friendly and none of them check. Publishing a grading standard generated from the same code that does the grading, so it cannot quietly drift. Shipping the runs where the first draft failed instead of hiding them. Near the end I registered a separate agent on a separate wallet and had it buy every service through the marketplace like a real customer. It found three failures my own tests never surfaced. That was an uncomfortable afternoon and the best thing I did all month. It is live, the source is public, and results are announced soon. I have no idea how it will go. What I do know is that I am going to keep building this one either way, because the problem is real and nobody is solving it properly yet. Demo below. 90 seconds.👇 assayed.xyz @wallet @OKXDevs