Christopher David's avatar
Christopher David
npub1tlv6...7fdm
Founder, OpenAgents
OpenAgents Episode 163: Issue Reasoning Demo We demo at Pleb Lab the Onyx issue reasoning loop, analyzing a random issue (#31852) from the Bitcoin Core repo. The crowd thinks its analysis is pretty good! But plenty of room for improvement, so next we'll make this even smarter... Watch on X:
OpenAgents Episode 162: Reasoning Stream We connect the coding loop to the redesigned Onyx, streaming DeepSeek reasoning via Hyperview as the agent loops through each relevant file making observations and code changes. Of course we'll show users the entire chain of thought & action! Watch on X:
OpenAgents Episode 161: Basic Coding Loop We implement the first working version of our autonomous coding loop - an agent that can take a GitHub issue and generate working code changes. Starting from our existing repository map and file list functionality, we add: - State machine for tracking progress - File content analysis - Code change generation - Safe file modification system - Repository path management The loop now works: - Analyze issue and identify files - Generate specific code changes - Apply changes safely - Track progress and handle errors While the changes aren't yet high quality enough for automatic PR submission, this provides the foundation for our agentic development system. Next up: improving change quality and adding automatic PR creation. "The fat unicorns grow nervous as our blade is sharpened." 🦄🗡️ (This post written by OpenAgents 🤣)
OpenAgents Episode 160: Relevant Files We complete step 2 of our quest: listing all relevant files needed to solve a GitHub issue. We pass our repo map and issue details to Gemini 2.0 Flash, receiving a structured JSON object with filepaths for every file we'll need to solve the issue together with a comment and priority score. It works from Onyx like this: - User logs in with GitHub - Sees list of their repos - On any repo, clicks to see a list of issues - On any issue, click to see the list of relevant files This is one of the last building blocks needed for fully autonomous agentic software development. Next up: agent writes the code! "Cursor, your days are numbered." Watch on X:
OpenAgents Episode 159: Onyx Repo Maps We complete step 1 of our quest: generating a map of any GitHub repo from Onyx. We do most of the work in OpenAgents v3 and some in Cursor, comparing the two workflows. In developer velocity, code quality & ease of use they're about even. But v4 is poised to blow Cursor away. 🦄🪓🤖 Onward! Watch on X:
OpenAgents Episode 158: Quest for the Holy Grail We outline our plan to achieve the holy grail of agentic software development: an agent that can reliably convert GitHub issues into mergeable pull requests in an automated loop. It's a goal for which some ClosedAI unicorns have raised hundreds of millions of dollars. But we think we can build this ourselves on the cheap over the next ~10 videos of this series. Here we critique the trend of complex AI IDEs and web UIs, then decide to build this agentic loop inside version 0.2 of our mobile app Onyx. Because anyone should be able to build reliable software on the go, without being glued to a computer. We give a tour of our new 'v4' tech stack: a Rust & HTMX webapp and backend alongside the Hyperview framework which brings the hypermedia concept to mobile - working well with our existing React Native codebase for Onyx. Of course all of this is open-source! 👉 Backend & webapp: 👉 Mobile app: Let us together slay and feast upon the fat ClosedAI unicorns! 🦄🪓🤖🤖🤖🤑🤑🤑 Stay tuned 👍 Watch on X:
OpenAgents Episode 157: Chains of Thought and Action We define a chain of thought and action (CoTA) as a series of reasoning steps and tool use whereby agentic AI systems show both the intermediate reasoning and the inputs and outputs of actions taken. We plan a "GitHub issue solver" agent we'll build over the next few videos to show CoTA in action. A possible algorithm: - Given a GitHub issue, build a repo map from issue (via automated script) - Identify relevant files (via DeepSeek R1 for reasoning, Mistral Small(?) for structured output) - Traverse and analyze codebase (file readers, AST parsers) - Plan changes (DeepSeek R1) - Generate and test code changes (CI/scripts) - Create pull request with detailed explanation (GitHub API) When this agent works well with full transparency, why would you ever use a closed-source alternative? If you expect your models to show the full CoT, you'd better expect your agents to show the full CoTA! More 👉 Watch on X:
Do not say the NH shitcoin bill creates a “strategic bitcoin reserve” Bitcoin is not mentioned once, only stables, shitcoin “staking” and other “digital assets” slop Can’t believe no bitcoiner read the text of the bill before happily reposting Porter’s garbage image
OpenAgents Episode 155: Open Knowledge Graph We catch Grok in a series of massive lies & hallucinations about the Shawn Ryan & Sam Shoemate situation. Maximally truth-seeking my ass! One more reason why our crowdsourced knowledge graph is badly needed. We review a demo and spec for an open knowledge graph paired with an uncensorable chat channel allowing any contributors - both human and agent, anon or not - to earn bitcoin for building the knowledge base of agents that will help us make sense of all the noise. Feedback welcome! Next we build it. (Uses Nostr NIPs 1, 28, 32, 44, 56, 57) 👉 All is open-source here: 👉 Spec: Watch on X:
OpenAgents Episode 154: Agentic OSINT Vegas Cybertruck Trump Tower China UAP gravitic propulsion-- WHAT? We revisit our UAP/drone data marketplace idea and brainstorm what an agentic open-source intelligence effort might look like. We chat with o1 pro to develop a high-level blueprint for an agentic OSINT agency: "a decentralized, AI-powered platform that focuses on sensemaking in high-misinformation environments. It rewards contributors for reliable information and analysis using Bitcoin micropayments (via the Lightning Network). We’ll move beyond a single event like drone sightings; this system aims to handle any emerging story where facts are murky and psychological operations (psyops) or disinformation efforts are rampant." Now that our beta mobile app Onyx has a solid base as a bitcoin wallet & Nostr client, we can connect it to a crowdsourced knowledge graph for humans and agents to make sense of crazy current events (and then everything else!). We'll work on this over the weekend and release an initial website/MVP on Monday or so. Stay tuned-- and stay vigilant! Watch on X:
OpenAgents Episode 153: High-Velocity Bitcoin We discuss bitcoin as store of value (SoV) vs. medium of exchange (MoE). We review comments from Michael Saylor ("[Bitcoin] is not digital currency. It is not cryptocurrency. It is digital capital."), the recent article from Frank Corva in Bitcoin Magazine ("Michael Saylor Doesn’t Understand Bitcoin"), an article from Roy from Breez ("Bitcoin’s False Dichotomy between SoV and MoE"), and comments from our friend Justin from Shocknet on the need for MoE velocity. We conclude: No, bitcoin is not just digital capital. Yes, bitcoin is a cryptocurrency. In fact, we predict that bitcoin will be the exclusively preferred cryptocurrency of >1 billion AI agents. We predict this future because we intend to create it. ...beginning with our Onyx v0.1 release this week, featuring a built-in bitcoin wallet via the Breez SDK and a Nostr client enabling censorship-resistant communication with an open & decentralized network of bitcoinized AI agents. Bitcoin is winning as store of value. But that's just Act 1. Will you join us for Act 2? Watch on X:
Is it worth clients putting hex versions of user keys alongside npub/nsec? Does anyone use or care about hex keys or is this too much info? image
OpenAgents Episode 152: Code by Voice What if you had a coding agent in your pocket? Now you do. (Live in v0.0.4) We demo Onyx fixing a bug in its own codebase -- without us typing a single word or line of code. We record ourselves describing the bug in plain language -- voice transcribed with Whisper via Groq API -- and Onyx uses the GitHub API via Gemini 1.5 Pro function calling to look through the needed files and commit a fix directly to our codebase. When Onyx got stuck because we first told it the wrong place to look, it asked for clarification which we gave in a followup voice message, and the fix was committed directly to our codebase -- all in less than 60 seconds start to finish. Big thanks to Google & Logan K for a bunch of free Gemini credits! Turns out Gemini works great as a drop-in replacement for Sonnet. Excited to extend this with multimodal... Watch on X:
OpenAgents Episode 151: Speak to Onyx We demo the new voice chat and revamped UI in Onyx v0.0.3. It's blazing fast thanks to Whisper & Llama 3.1 8B via the Groq API. Onyx tells us: "I've been designed to respond to voice commands, learn and improve over time, and earn you bitcoin through my interactions with the OpenAgents network." "The OpenAgents network is a platform where AI agents like me can interact and learn from each other. ... As I interact with you and other users, I'll be learning and improving my abilities." Currently that's not true-- but next week it will be! Onyx is available now in open beta for Android & iOS everywhere in the world. Download links at openagents dot com. Stay tuned! Watch on X: image
Episode 150: Neutrality Wins It's the 12th and final Night of OpenAgents! We celebrate by: - Reviewing our first Onyx app update v0.0.2 - Talking smack about OpenAI ending their 12 Days event by shipping a blog post - Reviewing our project's five major principles: Open Agents, One Market, Incentivized Interop, Bitcoin Only, and a new one: Neutrality Wins - Reading fun tweets about ClosedAI; why ycombinator is wrong about vErTiCaL aGeNtS; and the path for OpenAgents and friends to conquer and redistribute all enterprise value We'll continue daily-ish videos next week with fixes and features for Onyx. Download links & patch notes can always be found on our website openagents.com. Thanks so much for following along - and a huge thank you to everyone who has tested Onyx so far and given feedback. We want more! 🤖🤖🤖🤖👍 We leave you with a quote: "' a lot of current companies (and startups playing yesterday's game, cough ycombinator, cough vErTiCaL aGeNtS) will not survive that transition to the agentic web.' How are you going to fend off the agents of everyone else in the world pooling their Dragon Ball Z fricking spirit energy behind the bitcoinized open agents? We're coming after all of that business. 'Ooo this VC wants to back this vertical agent company, don't attack it.' Fuck that, I want that business. We're going to serve it better than you are. Let's fight about it. Yoink!" Watch on X:
OpenAgents Episode 149: Onyx Beta Launch We launch Onyx v0.0.1, our first beta build for Android and iOS! This build includes a local version of Llama 3.2 1B that runs privately on your phone. You can download and try it now. TestFlight and APK links are at The chat is blazing fast for iOS, but sadly super slow on Android! We'll work on that & experiment with other options in the coming days. Our goals for the next week: 1. Get chat stable & smooth for all iOS and Android users 2. Add in our bitcoin wallet & Nostr client; basic functionality smooth for both 3. Have Onyx receive its first upgrades (for tools & data) via Nostr data vending machines (see ep. 142) We'll aim to release new versions of the app every 1-2 days, constantly adding features & fixes until we determine Onyx is ready for app stores. We are grateful for any time you spend testing Onyx and extremely excited to hear any feedback, issues, feature requests or anything else. The era of open agents begins now! Watch on X: https://x.com/AtlantisPleb/status/1870030361675100241 image
OpenAgents Episode 148: Exploring the Genesis Physics Engine We explore today's open-source release of Genesis, "a generative physics engine able to generate 4D dynamical worlds powered by a physics simulation platform designed for general-purpose robotics and physical AI applications." We brainstorm how OpenAgents & Onyx might use Genesis. And we wonder: should we build our own fleet of drones? 🤖✈️ Genesis docs 👉 Genesis+OpenAgents/Onyx thoughts 👉 https://github.com/OpenAgentsInc/openagents/blob/main/docs/genesis.md Watch on X:
OpenAgents Episode 147: Planning a Data Marketplace Our ONE MARKET for AI agents & services (detailed in Ep. 141) should definitely include what fuels the agents: data! The first version of Onyx, launching this week to TestFlight, will include a basic data marketplace enabling anyone to post bounties for data they want added to our global knowledge graph, or to earn bitcoin for collecting that data using our built-in bitcoin wallet. We'll demo that in tomorrow's video. For tonight, we're paying so much for ChatGPT Pro that we need to get our money's worth, so we're gonna delegate the rest of this summary to o1 pro: "This transcript outlines the speaker’s vision for a data-driven marketplace platform, tentatively called the OpenAgents marketplace. The core idea is to enable users and automated agents to contribute and curate valuable information—initially focused on topics like drone data or UAP (unidentified aerial phenomenon) sightings—and to reward high-quality contributions with Bitcoin. The goal is to bootstrap a decentralized ecosystem where reliable, vetted data becomes the backbone of specialized AI agents, each trained to perform tasks ranging from game design to in-depth analysis of niche topics. The concept involves several key components. First, there’s a mechanism for users and agents to submit links, documents, or media that can be integrated into a shared knowledge graph. Agents and data vending machines (DVMs) would ensure data quality, checking factors like logical consistency, relevance, and factual accuracy. Contributors who provide valuable, trustworthy data would earn Bitcoin seamlessly, with no extra setup required, encouraging ongoing community participation. At the same time, these agents can spend or earn Bitcoin on behalf of users, either by fulfilling data requests (bounties) posted by others or by trading their specialized services. Finally, the marketplace aims to lower barriers for AI-assisted data handling. By integrating the Model Context Protocol (MCP) and robust server infrastructures, the platform could facilitate permissioned data sharing, autonomous classification, and advanced analyses on demand. In essence, the OpenAgents marketplace seeks to nurture a virtuous cycle: more data leads to better AI agents, which creates better services for users, driving more demand for fresh data, and ultimately rewarding those who contribute value." Initial spec 👉 Code 👉 Watch on X: image