Yohei Nakajima

this was my pypi hardening strategy for @activegraphai: - spin up a new @replit - point at docs page, ask it to build something - ask it to write a feedback report to package builder - feed that feedback into claude code - repeat

models underestimate how much work it takes (token usage) to accomplish a task, just like us

self-modification

i had to pack away my coding agents on monday cuz I knew I’d stay up too late if I played w activegraph during the week (yay, it’s Friday!) gautham kept playing with it and just showed me a custom UI 😁 I love open source

increasingly starting to feel like long-running agents need both a predictive world model and an experiential world model

make something agents want

other terms i’m learning: blackboard, DDD, CQRS, Akka, Kafka, BASE…

3 hour turn around on custom feature request!

okay i think this is a much better visualization of what i mean by "log-centric agent architecture"

RT sharan ⚡️ Yohei is probably the most prominent VC builder there is Humbled and honored that he's moving over to @cura_inc for his portfolio!! (Also, it's true he gives the best feedback out there)

RT Kunal Bhatia Superintelligence will be built on Self Improvement. Today @hexoai, we’re excited to release ‘SIA’ - an open-source Self-Improving AI, to achieve any goal through recursive self improvement. While trying to solve a problem, SIA doesn't just improve it's abilities by updating it's harness, it updates it's own weights as well.

two weeks ago i didn’t know what event sourcing was and now i see it everywhere

if you're a VC and want a killer portfolio monitoring tool like ours, but won't want to build it yourself... check out http://cura.inc

RT Sanchit monga “The future of AI is going to be local models running on extraordinary desktop hardware.” - @Jason This line from the recent @theallinpod hit hard. For years AI meant sending everything to the cloud, paying per request, and hoping for the best on latency and privacy. That era is ending. @Apple Silicon, AI PCs, and high-memory desktops are shifting the game. Local inference brings lower latency, zero marginal cost, real privacy, and actual user control. At @RunAnywhereAI, we're building exactly for this future: AI apps that run close to the user, understand private context, and work across devices without shipping your data to a third-party cloud. The next wave won't just be won by the biggest models. It will be won by the models that run where the user is. What workloads do you think go local first? #localai #inference #runanywhere #edgeai

RT Kat McGuire Re @yoheinakajima is genuinely the coolest power user you can have as an ai company got to talk to him the other day about how he's using cofounder and left thinking about how i coddle my agents too much. for ex., yohei literally gave cofounder a repo and told it to make him a marketing website and it made http://activegraph.ai i'm now a much more cruel overlord to my agents, and although i may be punished for it when the machines rise up, we're (me + cofounder) making some truly beautiful things rn <3

if you’re working on a side project that you might want to incorporate later, I highly suggest you start using http://cofounder.co

RT andrew pignanelli Yohei has the longest running session of any of our users and has been instrumental in our product development process. Super excited to see how activegraph grows!

RT General Intelligence Company "If I had Cofounder three years ago," says @yoheinakajima, "BabyAGI might have been a company." Our newest case study tracks our most active user, Yohei Nakajima, building @ActiveGraphAI - the event sourced graph runtime for long running agents.

self improving agents are here http://agnost.ai

this is cool. one paid api to wrap all paid apis

RT Garry Tan Great to see a GBrain x ActiveGraph crossover here Great retrieval pairs well with almost anything you want to do with agents

RT Payman your bank knows what you did. it has no idea why. that context, the why behind every transaction, is the most valuable data in banking. and nobody's capturing it. this changes with agentic banking.

RT Josef Chen Launching our new paper on arXiv: we trained the largest multilingual food model ever built. 4.1M recipes. 7 languages. 1,790 ingredients. 300 dimensions. All of human cooking compressed into 2 megabytes.

RT Nasdaq Exchange “So we needed our own sandbox. And that’s where we started building E2B.” In partnership with @Wing_VC's 2026 Enterprise Tech 30, @mlejva, Co-Founder & CEO of @E2B, joins Nasdaq to outline the infrastructure gap that gave rise to E2B and what microVM architecture means for the future of AI development. Watch the full interview here: http://spr.ly/6010B887T4

RT Artificial Intelligence Papers The Log is the Agent: Event-Sourced Reactive Graphs for Auditable, Forkable Agentic Systems Yohei Nakajima https://arxiv.org/abs/2605.21997 [𝚌𝚜.𝙰𝙸 𝚌𝚜.𝙼𝙰]

Healthspan just launched BioAge+, a biological age score that breaks down which systems are driving it and why kind of like a context graph for your health

we've been building agents around the llm, starting with conversations, adding tools, giving rules, logging everything, and storing a form of it as retrievable state Active Graphs flips this. what if we treated the logs as the agent, which include rules provided (and changed), tools given/used, and conversations. behaviors (including llm calls) sit on top of this, react to changes, and then create changes back into the state. this is more like how we work. we don't have workflows, just a whole bunch of rule based behaviors that sometimes chain. we are also not defined by our reasoning capability, but by our beliefs and experiences that shaped them.

RT Bill Lennon there he goes again, @yoheinakajima blazing the paths to continuously running self aware agents. this and the posts he links are super worth reading.

RT Jon Radoff 👾/acc 🎮 Metavert LLMs are stateless (every time you reply to an LLM, you re-inject the entire conversation to a fresh inference) the purpose of memory is to provide continuity (in games we'd call it a 'persistent world') that optimizes the context so the LLM can efficiently answer questions about the world LLMs inverted a lot relative to software patterns: most software acted upon the world; with LLMs we have to construct the world around the LLM maybe the memory layer becomes the primary artifact (the persistent world), and LLMs become interchangeable reasoning engines plugged into it. the 'model' gets swapped out, but the agent's identity and history persist

RT Yohei Re @ElevenLabs here's the tl;dr, thanks @grok

if this works, will be great for self-improving agents because of the ability to fork and diff agent runs

this was not on my bingo card

what if we mapped older distributed systems patterns (like actor models or reactive, state-driven blackboards) to LLM agents?

RT Ariel Jalali ⚙️💬 The daddy of baby AGI articulates what we all feel is missing in the current state of AI architecture

I didn’t really understand what people meant by stateful agents so I started exploring my current take is that we’re not there yet, which led to an interesting experiment…