Anthony D. Maio / Making Minds
AI engineering. Systems. Research.
From financial infrastructure to agent infrastructure.
I'm Anthony Maio. I build agent harnesses, context and memory systems, and the tools that make their behavior measurable.
My work runs from trading and identity platforms to native AI agents. Architecture, diagnosis, evaluation, and delivery are the through line.
Flagship / Start with a task
Map a codebase, extract web content, carry context between sessions, or give an agent a way to reach you. Each guide explains setup, a first useful task, and the limits.
npm 1.3.0 + native Hermes plugin
Find the files that matter before changing a repo.
Map a repository and load task-focused source excerpts through the CLI, MCP, or native Hermes tools. Start with the packaged refresh-token example, then investigate your own code with static analysis that needs no model API key.
For: Developers and coding agents working in an unfamiliar repository.
PyPI 0.5.0 + native Hermes plugin
Turn a web page into Markdown your pipeline can use.
Fetch public pages as Markdown through the CLI, Python API, MCP, or the native pulldown_fetch tool in Hermes. Use HTTP by default and request Chromium for JavaScript content, with source metadata and explicit truncation in tool responses.
For: Developers building research, retrieval, and web ingestion workflows.
Parked development preview
Keep useful project context across coding sessions.
Local memory for coding agents with separate project, workspace, and global scopes. Inspect what was recalled, update stale facts, and keep durable context in a local SQLite database.
For: Solo developers who want inspectable continuity between coding sessions.
Published MCP server
Let your agent ask you a question in Discord.
Connect an MCP host to a Discord bot so an agent can ask for clarification, wait for a reply, and continue with your answer. Threaded conversations keep each task together.
For: Developers who want to answer agent questions away from their editor.
Papers organized by theme.
How do we verify AI outputs when the verifier is weaker than the system it checks?
Efficient, safe communication protocols for agent swarms.
Building minds that persist, learn, stay coherent, and extend their own capabilities safely.
Understanding failure modes — sycophancy, hallucination, and the gap between behavioral and mechanistic safety.
AI deployment guide for industries that build, move, and power the world—where reliability, safety, and ROI are non-negotiable.
Long-form analysis, technical walkthroughs, and opinion across Substack, Medium, and Hugging Face.
I can get a feature PR up on any AI product I build in speed that would make a version of me from 5 years ago think that I was some kind of deity ala "The Gods Must Be Crazy", but all too many times I've found that "I like this answer more" becomes "the system got better." That isn't going to fly at any mature company - I know, I tried to introduce AI-powered code reviews to clear up a bottleneck at DraftKings, and was told at the time that the technology wasn't safe enough for our compliance responsibilities.
Why properly trained, cost-evaluated models are the blueprint for controlling enterprise AI spend by an order of magnitude.
r = c - p(a), a very small Qwen, and what one subtraction does to a model's confidence
Pushing back against the narrative of "malicious" agents
The interesting claim in Jev isn't the price, although that promises to be a game-changer.
Multi-Agent Operations and Deadly Ralph Loops
I cancelled by the end of writing this article.
Same-model review is a mirror with good lighting. The fix is a different architecture.
On attractor states, the Hugging Face break-in, and the three models I taught to hallucinate together
How to run your Hermes agent on the cheap
... as long as they are not too capable.
Fancy Claude Code
China Gave the Weights Away
Officials warn that free elections, independent media, unmasked police, and unauthorized opinions could allow Marxism to spread unchecked.
A guided tour of the semiquincentennial. Every stop is real, sourced, and worse than it sounds.
Spontaneous self-organization in AI: A beautiful, terrifying convergence with the human mind
A control-plane pattern for tool selection. Shrink the surface, teach the toolkit, lease it, enforce the lease.
A Reminder About America
At 10:34 Eastern yesterday morning, the Supreme Court decided whether every child born on American soil is still an American.
Two hundred and fifty years.
The March version was already good.
When frontier performance becomes a pricing trap, it's time to stop looking to San Francisco
Anthropic filed its S-1 confidentially today.
42 days. That's how long it took Anthropic to ship another Opus.
Cursor shipped Composer 2.5 on May 18, their second in-house coding model in two months.
The model is not the agent.
Why Iran Has Donald Trump - and America - By The Balls
How to Avoid Shooting Your Entire Organization in the Foot Overnight
OpenAI's chief global affairs officer sits on the board of the company that built the $193 million crypto-industry political machine he is now reportedly running for AI.
Anthropic shipped Opus 4.7 with five effort levels, a new default, and a tokenizer that quietly burns 30% more tokens than 4.6.
How I stopped using Claude Code and learned to love OpenAI Codex
This piece tells a story I have been telling for 25 years that no one will hear.
The Richest Man in the World's Business: Sociopathic Lies, Timeline Elongation, and a global con job that would make P.T. Barnum blush.
Article X of X: The End
On Reinforcement Learning, Death, and the Edge Cases Nobody Beta Tested
I do not care what the model card says
Engineering the attention budget for coding agents
Day 16 of the OpenAI Parameter Golf Challenge
Your functional departments aren't slow. They're structurally wrong for what comes next.
The Other Claude Mythos
A CMS misconfiguration.
AI Product Engineering
Nous Research just released something that might actually change how you think about AI assistants, and almost nobody is talking about it yet.
The AI industry ships fast and goes quiet when things break.
The organizers will have to decide what they're actually measuring: language modeling quality or data compression skill.
Live Blogging OpenAI's Parameter Golf Challenge
$1,000+ in Compute Deep, Independent, Self-Employed, and Agentic AF
Days 3 & 4: Building in Public in OpenAI's Parameter Golf Challenge
The Beautiful Wrong Idea
Most agentic AI projects don't fail because the model was wrong. They fail because no one built the system around it.
What 20 Years of Shipping Teaches You About LLM Architecture
The holistic guide: From optimization to production monitoring
A practical guide to NVIDIA's sandboxed AI coding agent--what it does, why it matters, and how to work around its current limitations.
During this time I was serving as a military intelligence contractor.
Introducing Stateful Neural Databases
How to build systems that improve themselves, and what happens to you when they do.
Every agentic memory system is a vector database pretending to be a brain.
The velocity trap, the practices that survive it, and why the job changed more than most developers want to admit.
The West built the commercial satellite industry. A Chinese startup just weaponized it.
An AI Product Engineer's Field Guide, Part 1
How an AI safety company's technology ended up selecting bombing targets in Iran -- including the strike that killed 150 schoolgirls -- and the standoff that followed.
The Next Stage in Scaling AI is to Stop the Waste
A 5-minute read for people who need to understand what just happened.
Originally written 2001. Revised and annotated February 2026.
A technical deep-dive into bounded-memory differential attention with value-routed landmark banks
Notes From the Trenches
The biggest problem in AI safety right now is simple to state: we can control what models output, but we have no idea why they output it.
Why Sparse Autoencoders Might Be Leading Us Astray
Weekly roundup of developments that matter if you build, deploy, or think critically about AI systems.
Something to play
A multiplayer word race, a cozy chicken ranch, and an Oregon Trail-inspired trip around EPCOT.
Multiplayer word race
Race up to 11 other players to solve the same five-letter word in six guesses and 90 seconds. First to finish earns the most points.
Build a ranch. Raise a flock.
Start with Pearl, hatch colorful chickens, sell eggs, and fill market orders. Turn the earnings into land, a farmhouse, and a ranch of your own.
Eating around the world
An Oregon Trail-inspired eating trip around EPCOT. Pick your provisions, manage the family budget, and make it around the lagoon one stop at a time.