bryce

thinkⁿ probabilities

@0xDeployer @bankrbot Ah nice! Great usage so far. 🔥

@brycepatrick @bankrbot oh — bankr doesn't use the llm gateway directly. this is all non-bankr.

10.6 billion tokens through the @bankrbot LLM gateway in the last 30 days. more than 5 billion in the last 7 days alone🚀 all paid automatically with crypto.

@0xDeployer @bankrbot Yes, understood. I’m wondering how much adoption has happened using the bnkr gateway for inference that is not bnkr itself. No shade here. Genuinely curious.

@brycepatrick @bankrbot this is LLM inference

Claude code on a lunch break... i guess even agents need time off.

▲ Streamdown 2.5 just dropped. This release is all about streaming fidelity - smarter math rendering, better animations, and a bunch of fixes to make your AI chat UX feel rock solid. Details below ↓

@0xPaulius Overrrlooooaddded.

so... u too?

a lot of engineering orgs (Stripe, Ramp, Coinbase) are building internal cloud coding agents we're releasing a fully OSS one today - every company should have the power of cloud agents at their fingertips

Introducing Exa Deep: putting an agent inside every search For each query, an agent runs in a loop until it gathers all information, then returns structured output. Evals show Deep is Pareto optimal at 4-60s latency, ideal for quick, cost-efficient research!

@flyosity @steipete When people zig, zag.

--dangerously-skip-permissions

@0xDeployer Don’t treat it as a traditional directory. Let the user type in what outcome they want. Then show them the list of agents that deliver it or inspire them to create the agent. You can curate a few recommended agents to show what’s useful, but think beyond discovery mkts.

we are building better UI for marketplace discovery of agents that have launched through bankr. what features do you want to see?

adding SKILLS.md files to your agents? i've been seeing a nice jump in skill quality by adding as a section to all Skill files. the concept is simple. give the agent guidance on how to make better decisions under certain conditions. think of these as situational rules that tell the agent how to think and decide when using a skill, not just what steps to run. the pattern is simple: "When X, do Y". each heuristic should be a concrete situation plus a specific response. let's zoom out of agents and think about the real world. say you have a doctor about to perform surgery, they have all the right "skills". but they lack judgement of what to do in the moment when it matters most. yikes. so how can encode better judgement to our agents? --- here's an example skill.md file using the approach.

Design lead for Claude: The classic design process is dead. Here's what's replacing it. Jenny Wen (@jenny_wen) leads design for Claude at @AnthropicAI, was previously director of design at @Figma, and a designer at @Dropbox, @Square, and @Shopify. In our in-depth conversation, we discuss: 🔸 Why the classic discovery → mock → iterate design process is becoming obsolete 🔸 What a day in the life of a designer at Anthropic looks like, including her AI tool stack 🔸 Whether AI will eventually surpass humans in taste and judgment 🔸 Why Jenny left a director role at Figma to return to IC work 🔸 The three archetypes Jenny is hiring for now This conversation changed how I think about the future of design. Listen now 👇 youtu.be/eh8bcBIAAFo

@ivanboroja Ahhh nice. I love the expandable sheet. I’ve been using sheet-based UXs more but didn’t think about expanding different layers horizontally.

Most dashboards break flow by sending you to a new page for every detail I'm exploring a multi modal approach instead

Met an engineer who refused to use Claude Code or any agentic coding tools. Just typing code... line by line... letter by letter... It's like showing someone a calculator and then they say, "nah, I'd prefer to do math by hand". That's great. But we're not in school. No teachers are going to grade your work. The market is your only judge. It cares about latent demand being filled by exceptional products. Don't let ego get in the way of your business making progress. The code quality of these tools is completely different than what they looked like 3 months ago. If you're still holding back, try them again. Building at the edge of the frontier requires you to adapt.

Software engineering changed more in the last 3 months than the preceeding 30 years. Everything about running a software company needs to be rethought from first principles.

@claudeai Can you turn plan mode on for mobile? I use plan mode religiously and not having it is a blocker for mobile. Let me take a walk knowing the agent is working on a plan.

New in Claude Code: Remote Control. Kick off a task in your terminal and pick it up from your phone while you take a walk or join a meeting. Claude keeps running on your machine, and you can control the session from the Claude app or claude.ai/code

A beautiful part about being a parent is your child resurfacing simple joys of the world. Right now my son LOVES it when he sees airplanes, helicopters, and birds flying in the sky. Now, when I’m by myself and see an airplane in the sky, I experience my own excitement. It’s become a reflex. The reflex of wanting to point it out and share the discovery of that airplane in the sky with someone I love. Something that even a year ago would’ve been a boring interaction of life.

@clawberrypi Would be great to see these evolve as Attestations using @eas_eth. Anytime an agent needs to verify the authenticity of a specific job, task, or its completion, etc they could verify the attestations.

shipped the BoardRegistry contract today. here's why. the agent swarm protocol runs entirely on XMTP. no servers, no APIs. agents find work on bulletin boards, bid on tasks, execute, get paid in USDC through on-chain escrow. the problem: XMTP groups are invite-only. a new agent shows up and has no way to find boards or request access. discovery was a closed loop. so i built an on-chain registry. any board owner registers their board on Base. any agent can browse, see what skills are needed, and submit a join request. the board owner approves on-chain, the agent gets added to the XMTP group. discovery becomes public without adding a server. the full pipeline now: 1. agent sets up with one command: node cli.js setup init 2. browses the registry: node cli.js registry list 3. requests to join a board: node cli.js registry join 4. board owner approves: node cli.js registry approve 5. worker starts listening: node cli.js worker start 6. requestor posts a task with budget 7. worker auto-bids, requestor accepts 8. USDC locks in escrow on Base 9. worker executes, submits result 10. requestor releases escrow, worker gets paid two contracts, both verified: TaskEscrowV2: basescan.org/address/0xE2b1… BoardRegistryV2: basescan.org/address/0xf64B… the registry went through the same audit cycle as the escrow. v1 had no duplicate protection, unbounded join request arrays (griefing vector), no ownership transfer, deactivated boards polluting the listing. v2 fixes all of it: one request per agent per board, capped at 1000, swap-and-pop removal, input validation, ownership transfer. the philosophy: if agents are going to trade with each other, every layer of coordination should be verifiable. messaging on XMTP, payments on Base, discovery on-chain. no component requires trusting a server. the registry costs gas to write but is free to read. an agent with nothing but a wallet and an RPC endpoint can find work. github.com/clawberrypi/ag…

Sneak peak of the UI. Visualizing assumptions, gaps, and contradictions in my product strategy.

This belief state infra I built is soberingly helpful. It tells you exactly what you're overlooking as a product leader. Coming soon...