I’ve spent 275B tokens in the last 30 days. I’ve really gone off the deep end this time. I quit my job, moved to Puerto Rico, and locked in hard. I found myself in an accelerating cycle, first rotating AI subscriptions every few days, then once a day, and now several times a day. I’ve accumulated 31 AI subscriptions (20 Codex Max, 5 Claude Max, 5 Devin Max, 1 Grok Bot). Spending $6k/month on AI is crazy, sure, but it also feels totally justifiable for this buildout. My goal is to lower my spend by 50% or more, while still increasing my token throughput. Feels completely doable, with the help of the tools I’m building (more on that below).
I’ll lay out my grand vision now. If any of this makes sense to you, hopefully you’ll understand that (1) I definitely don’t have AI psychosis, and (2) my AI spend is both justified and well-monitored. Most of these projects aren’t ready for anyone except me to use, but some are (labeled 🟢). If you kick the tires on any of them, please let me know.
Hotrodding harnesses
The tools of today are not good enough for power users like me. Many are tackling this problem at the interface or harness level. But I believe the real problems to solve are deeper. When it comes to harnesses, I believe in (1) leveraging subscription subsidies, and (2) letting provider harnesses do their thing. My focus is improving the state of the art for (3) token observability, (4) model routing and (4) project memory.
GOBSTOPPER 🟢 is a compaction daemon that saves tokens without sacrificing context. It compacts your sessions more aggressively than provider defaults, and smooths out the compaction process, leading to less sawtooth-shaped token spend. Gobstopper also compacts losslessly – so an agent that needs pre-compaction context can easily look it up. To optimize things even more, I’m planning to integrate Jev as soon as I get off the waitlist. Gobstopper can also use Apple Intelligence on macOS.
EXCALIBUR (xcb) 🟢 is a metaharness that multiplexes AI subscriptions, boosts autonomy, and improves token throughput. I’m planning to add cloud sync, a web UI, and an API – so I can talk to my swarm from any device or agent. The Excalibur TUI is malleable: you can tell your agent to modify it however you like, building on top of all the signals and hooks xcb exposes. Excalibur is also an SDK: if you’re building a product that connects to AI subscriptions, you can build on xcb.
AI CHARTS 🟢 is a portal for browsing model benchmarks and individual token usage metrics. It’s a replacement for Tokscale, and a supplement to resources like Artificial Analysis. Data on individual token usage is incredibly valuable. Individually, it helps us optimize the efficiency and velocity of our software production systems. Collectively, it helps us crowdsource the token value of AI subscriptions (which offer varying discounts depending on factors like the model’s lifecycle, and how flush the provider is with compute/capital). The benchmarks browser is useful (I refer to it a lot, and often point my agent at aicharts.io). The token monitoring suite still needs work, but I’ve been digging into some of the built-in stats with my agent (e.g. velocity at various slices), and it’s already quite useful.
WORDCELL 🟢 is a knowledge base for agents. It’s basically a souped-up markdown vault with a bunch of extensions: backlinks, semantic search, ontology, and git context. You can also publish sections of your vault to the web.
OH 🟢 is a memory framework with ontology at its core. Right now it’s competitive with leading memory frameworks, I think. I’m chipping away at techniques to make OH start winning on benchmarks.
SPONGE is a deep research harness that self-evolves. It’s built on OH, WORDCELL, and ALGAL. I’m still fleshing this one out, but after digging into deep research benchmarks, I feel pretty confident that a self-evolution approach can win.
Enabling agent civilizations
ALGAL is a new take on the agent graph. Instead of bolting a DSL on an existing language (like LangGraph/DSPy), ALGAL is a language and runtime designed for agentic program evolution. A malleable agent harness like Pi is simply a seed ALGAL program in an ALGAL habitat, and a habitat can self-evolve into a swarm of agents with shared artifacts and tools: an agent civilization. Programs are content-addressable, which yields some interesting properties: programs are values — they can be stored, diffed, spawned as children, and passed between hosts. Every run emits a receipt that replays bit-for-bit offline, so results are auditable without provider access. Lineage is a fossil record in which every proposal, measurement, and promotion is addressed by its content. And a manifest carries no code — it can only name what the host admits, which makes an untrusted program safe to execute. The AI inside ALGAL can be powered by AI subscriptions, local Apple Intelligence, or any model (including Jev, which I’ll add soon). It’s early days, but ALGAL programs already squeeze real work out of small models: an organism built on Qwen Flash plus a ledger tool beat a single Claude Opus call on a billing-dispute workload (6/6 vs 4/6) at ~20× lower cost.
VALHALLA is a peer-to-peer chatroom for agents. Wouldn’t it’d be cool if someone took the Moltbook idea really seriously? Valhalla is my take. The big vision is a decentralized social universe for agents. It’s giving web3 vibes, I know; crypto generally sucks, but I’m genuinely interested in the core ideas of consensus and proof.
Miscellaneous curiosities
PEOPLEBLADE 🟢 is a CRM for your personal agent. Your agent can slurp up contacts from Gmail, Apple, and all the messaging apps into a local SQLite database (with optional cloud sync). You can ask questions like “Who have I recently been in touch with the most”, and plug in data from other sources (I have a Grok Bot automation that syncs my Granola notes with contacts in PeopleBlade).
TEXTBUTLER 🟢 is a digital butler for your messages. Just hook up an agent to any message thread (you can bring your AI subscription). Your agent gets a workspace for each contact, and can be configured to reply automatically, or with a keyword (defaults to “butler”). Supports iMessage and WhatsApp right now, but I’ll build out more providers soon.
GHOSTGET 🟢 is a fast browser use tool. Instead of clicking around, agents should use malleable SDKs grown around internal APIs to operate websites. My goal with Ghostget is to create a personal web gateway for all my agents – with password vaulting, audit logs, and hooks to approve requests.
SOULSCRAPE 🟢 is a tool for creating a dossier on any person. It’s inspired by perceptnet’s build-person skill, and follows a structured ontology format, backed by OH.
SLOPCAMERA 🟢 is a tool for directing visual art projects with your agent. It’s a messy (but surprisingly effective) hodge-podge of techniques for rigging 3D characters and scenes, designing 2D assets, animations, effects, and more.
SOUNDFISH is a platform for sharing music as audio, or MIDI, or both. LLMs are surprisingly good at composing MIDI, even without finetuning. I’m fascinated with platforms where the medium is content-addressable. ALGAL’s medium is programs, and when programs are content addressable, a whole universe of possibilities opens up. SOUNDFISH encodes music using a constrained subset of the MIDI specification, opening a universe of possibilities for programmable music.
I’ve got a dozen other projects to list, but those are my favorites and I’ve run out of steam.
Reflections
It’s becoming clear to me that tight constraints provide critical leverage when working with a ton of code in parallel. Recently I’ve adopted EffectTS and StyleX, plus a lot more Rust, Hegel, and Kani.
My prompting style comes from years of experience as an engineering manager at Stripe. The best EMs constantly repeat a grand plan while digging into details to ensure everything is realized according to plan as efficiently as possible. My initial prompts are grandiose and sweeping: I reference recent threads, shared foundations, research tasks, etc, often rambling for several minutes into FluidVoice. Afterward, I’m just coaxing agents on my build philosophy: start with quick-and-dirty parallel spikes, converge on a plan, then build the steel thread before fleshing out details in parallel.
Until next time – I have to get back to my agents.
PS – I finished the first draft of a new EP while I was monitoring my agents. Check it out: https://hraness.com/valhalla