Today we're launching B3OS in public beta. Onchain workflows any agent or human can run. Live today: 𓏠 Traders copy-trading on Polymarket or trading X headlines 𓏠 Developers powering onchain apps with scheduled sends, recurring swaps, and more 𓏠 Teams paying vendors in USDC from email or Slack Build in our UI or through our MCP server. This is the future of autonomous onchain finance.

You guys no idea how badly I want to tell everyone about what we are working on with @b3dotfun

Where does your data go every time you prompt a hosted model? B3 CEO @darylX24 on why that answer is starting to matter, and why local inference is becoming the default.

Why doesn't a single AI agent understand crypto the way it understands everything else? We've spent the year making B3OS the most natural crypto agent you can use. The hard part was never the reasoning, it was the execution: landing onchain actions reliably, keeping audit trails, and making sure that when Caddie fires a transaction or pings one of our 1,000 connectors like Slack or Telegram, it happens the way you asked instead of the way a hallucinating model guessed. Our CTO @seangeng on where Caddie goes from here:

In @PanteraCapital's The Convergence of AI and Blockchain, @cosmo_jiang makes the case that AI and blockchain push each other forward. Blockchain gives AI open systems to pull from. In return, AI abstracts away the complexity that has held crypto back. Abstracting it away is the hard part. The rails exist and the models can reason, but the layer that turns intent into reliable execution onchain is still early.

We're launching a solution for this soon.

Accurate sentiment analysis requires top models. Caddie AI in B3OS is live and using Claude Fable 5.

Where does crypto fit once AI is in everything? @darylX24's take on @latenightonbase: the models keep getting better, but someone has to turn that power into tools regular people can actually use. That's where B3OS comes into the picture.

Meet B3 (@b3dotfun) the operating system for onchain automation built by the former minds behind Coinbase Wallet, Base, and USDC. 🛠️ They just dropped B3OS (b3os.org), a modular execution layer that changes how we interact with crypto. Here is why it’s game changer: 🔹 No Code Workflows: Build complex trading systems & automations using simple Lego blocks. 🔹 AI Agent Ready: High reliability with low latency perfect for autonomous onchain tasks. 🔹 Zero Friction: Solves gas spikes RPC failures and chain reorgs out of the box. Backed by an ecosystem that includes B3 Labs & Strategic Holdings to drive long term value The future of Base is automated. ⚡️ #Base #B3OS #Web3 #AIAgents cc: @jessepollak @base @BuildOnBase @brian_armstrong @seangeng @ItsGioLogist @TimmyGruber @Gabedotninja @b3labs_

B3OS is an execution layer to instantly access 1,000 audited, battle-tested crypto triggers and actions (like @Morpho, @HyperliquidX, @Polymarket, @CoinbaseDev and more)

Great variety of guests heading to the stream this week • Former Base employees discussing the fallout from the recent layoffs • Coinbase-backed teams building the next generation of products • Founders operating in the AI agent space • Crypto researchers and industry personalities Each brings a unique perspective, story, and set of lessons worth sharing.

Big fan of agents interfaces being in native messaging channels like iMessage, telegram, WhatsApp My team is working on similar surfaces and it’s a super interesting engineering / product challenge to make the agents conversational / feel “natural” iMessage is especially a challenge cause you don’t have rich input elements like slack does with blocks or telegram / WhatsApp can do, so you need to come up with hacky (but still low latency) ways to present information to the user. For teams who use agents, you want them to behave like teammates and the problem is different shape in multi-tenant systems like slack or discord where you need to juggle user context, org context, and permissions. Two very different problems, same goal: make the agent feel like a person, not software you're operating. Gonna write up a blog post about it below

When you pay for inference today, you're renting it. ChatGPT, Claude, whatever you run, it lives on someone else's machine, metered behind a monthly fee. Our CTO @seangeng sees that shifting. Open models are catching up to closed ones, demand for compute already outruns supply, and inference is moving local, onto machines people own and networks they control.

Local AI is starting to feel more practical love their tagline, LM Studio’s new app lets you run local models, in your pocket when Android and who else is building surfaces like this? I want to try all of them

Set-and-forget only works if you trust what the agent is reading before it moves. That's the part @DatalineAI handles, and it sits right in front of where B3OS takes over to execute onchain. Glad to be a Dataline launch partner.

Been looking into token optimization and model routing, I think super obvious optimization to tackle both cost demand on inference Here’s a small post about different techniques and methods seangeng.com/writing/the-hon…

How long does it take you to pay all of your contractors in @USDC each month? This took one sentence.

In trading, stale data is just a slower way to be wrong. - Caddie reads the market live the second you ask. ETF flows, liquidations, whale moves, fear & greed, all current, synthesized from every reliable source. - But reading is just the start. same chat, you act on it: set a price alert, build a workflow, swap, hedge, automate the whole thing. The live data wires straight into onchain actions, wallets and gas already first-class on b3os.org. - General models are great at a lot of things. Staying live on the market is a different matter, and Caddie is built for it. right tool, right job.

Figuring out what a user actually wants when they're talking to your LLM agent is more of a PITA than you can imagine. We spent months on this with Caddie, our AI workflow builder at @b3dotfun. Started with regex. Ended up with a 6-phase hybrid classifier: 1. Pending plan? (deterministic) 2. Last LLM suggested a plan or asked a question? (deterministic) 3. User asked a question with no workflow intent? (deterministic) 4. Editing an existing workflow? (deterministic) 5. Domain keywords or action verbs? (deterministic) 6. None matched? LLM fallback. Our deterministic guards capture ~95% of intent. No latency, no tokens burned. The LLM only fires when we genuinely can't tell. Here's a visual. PS: Curious how others have handled intent classification. I'd imagine it's usually a hybrid. What LLMs downstream are you finding most performant for this specific use case?

Who owns/uses something like this currently? What do you use it for? I’m bullish on local inference and localmaxxing but prefer the laptop format