From day one of Brevy, the thesis has always been to utilize AI agents to achieve 100x more than a non-AI company. As we've been building towards that in the last several months, several hairy problems emerged.
- Work orchestration.How do we create the tasks that the AI agents should work on? How do we manage our human team?
- Task granularity.How large or small should the tasks be? How do we manage a huge number of task types in the future?
- Teaching AI agents.How do we build an AI agent that learns how to do every task that we've created? Building automations by writing code is painful.
- Agent scheduling.How often should these AI agents run? How do we prevent them running too frequently and clashing with each other, while not running them so infrequently that they're operating on stale state?
- Guardrails and interpretability.How do we understand why an AI agent made a certain decision? How do we prevent it from causing damage?
- Recursive self improvement.How can we get the AI agents to improve over time?
These are all fields of active research. There are several historic analogies and recent research breakthroughs that make me think that there is one simple answer to all of these problems.
The answer: remove the scaffolding. Unhobble the agents.
Historic Examples
In self driving, language translation, and voice AI there are massive examples of huge quality improvements once the creators got rid of their deterministic systems and went full AI.
Self Driving
I like using the analogy of autonomous vehicles. In the last couple years, both Tesla and Waymo have gotten insanely good. The reason? Neural networks, end to end.
Although Tesla and Waymo have very different technical approaches on autonomy, they both realized that they need to remove as much of their hard coded logic as possible and rely on larger unified AI systems. Now my car drives itself, and I don't take Ubers in SF.
Language Translation
Historically, around 2017, if you wanted to translate from one language to another, Google had to train a bespoke model for each language pairing. English to Spanish was different from Spanish to Arabic, which was different from Arabic to English.
The transformer, which is the core research unlock that brought us LLMs, was actually invented by Google to solve this language translation problem.
In short, they were able to remove all of the complexity and just have one model that could translate almost any language to almost any other language.
The result? Language translation is now absolutely insane.
Voice as a Modality
Up until a couple years ago, speaking to an AI model wasn't a great experience at all. You had to take speech and convert it to text, then take that text and run it through an LLM. The LLM then responds in text. You then need to convert that text into speech and send it to the user. This technically works but presents a bunch of problems.
- How do you know when the user is done talking?
- What if the user starts talking during the cycle I mentioned above?
- This means you can't "talk over" each other like what happens in natural conversation.
The whole system was very clunky. Then things felt like they suddenly got much better. You could talk to ChatGPT and it felt crazy, like talking to a human inside your phone. The reason is that OpenAI trained a voice to voice model. Rather than the series of steps mentioned above, the voice of the speaker is sent to an LLM directly and that LLM generates an audio response.
You can actually notice this today if you compare the audio mode of ChatGPT Realtime with the voice mode of the Claude app. Claude is garbage in this regard.
Modern Breakthroughs: LLMs Are Eating Traditional Code
Software 3.0
Andrej Karpathy was a professor at Stanford, then Head of AI at Tesla, and later one of the founders of OpenAI. I consider Andrej Karpathy to be AI God. When he speaks, we listen. He says that we are entering a new world of Software 3.0.
In short, what that means is that problems that required hard coded solutions are now going to be solved with AI agents.
Software 1.0 is the static code that we know and love. You hard code everything.
Software 2.0 are the niche machine learning models of the last decade. This is the custom English to Spanish model that Google trained, or the custom speech to text model for voice AI.
Software 3.0 is where the AI agents solve the problem end to end.
You hard code everything.
A bespoke model per problem.
The agent solves it end to end.
Reasoning Models
Around January 2025, the world was shocked by DeepSeek R1. It was partially due to a crazy claim that the Chinese were able to train it on a shoestring budget. What really made people in the technical community excited though, was that it was the first time we got to see the inside of a reasoning model. OpenAI had produced the first reasoning model, o1, but the code wasn't public.
What's cool about a reasoning model is that it basically allows an AI model to think and talk to itself. Before this, models like GPT-4 were "one shot" models. This means you ask them a question, and they just start responding one word after the other without the ability to reconsider, delete, change wording, and so on.
Reasoning models allowed the agent to take their initial response, reread it, think about it, rework it. For example, if you asked GPT-4 the derivative of cos(), it would have to answer without thinking, from memory almost. If you asked o1, it could think through the problem and derive the answer.
The result? You saw a huge increase in the capabilities of AI models.
Agents
Now fast forward a year and we had another massive breakthrough. AI agent capabilities went through the roof. In late November 2025 I was playing with Goose. It was supposed to be the AI agent framework that we use. Then during the Christmas break, Anthropic released a banger of a model, Opus 4.5, that caused everyone, including myself, to neglect their families and lock themselves in a room to code.
Soon after, Clawbot, a play on Claude, took the world by storm and Mac Minis around the world were in short supply. OpenAI bought that company real quick, for billions. Anthropic, getting a little jealous, decided it was going to go after entire industries like security, finance, and logistics. They made so much improvement in Claude Code that they ushered in the SaaSpocalypse. Entire industries saw their stocks collapse.
So what's causing all of this havoc? We've finally made great progress on the agent harness, the framework and software for how to string many reasoning models together to get the job done.
These massive improvements solved for us the six or more problems I mentioned at the top of the article.
This is why I said that the answer to our agentic AI problems is to remove the scaffolding and unleash the agents.
How Do We Unleash the Agents?
In order to fully utilize AI agents, we, and every other company in the economy, need to completely reimagine the way they operate. In the coming articles, I'll go into several topics around how we do this. I'll give a little teaser now, with full articles on each to come soon. Let's get it.
The Cyborgs Are Here: The Personified Agent
I think of this new era of AI agents as cyborgs. They're a new species, half human half machine. Conceptually, whenever we think of how to work with or build AI agents, we should think about how their fully human brethren would be treated. These cyborgs deserve names, personalities, and the rest of it.
At the same time, they have superpowers. They have infinite memory, can learn very quickly, naturally improve every six months, can work 24/7, and don't get upset with harsh feedback.
Everyone's a Manager: The AI Native Org
Cyborgs necessitate a change in company org charts. All human employees are now strictly in a managerial position. All actual work must be done by AI agents. Managers monitor performance, give training and coaching, and set direction. Managers don't do the work.
Misaligned team. Four people pulling in different directions. The resultant is nearly nothing.
Misaligned team. Four people pulling in different directions. The resultant is nearly nothing. Aligned team. The same four, pointing the same way. The vectors compound. Aligned, plus agents. Two people directing twenty agents, all parallel and much longer. The vectors explode: a 100x multiplier.Four people pulling in different directions, the same four aligned, and then two people directing twenty agents. It plays once on its own; after that, drag the slider or jump to a stage.
Simplicity Scales: The Role of Mission Control
Given we have new cyborgs that can do the work, Mission Control, our internal operations system, should be as simple as possible. It should be two things. One, a system of record to store the state of a given case. Two, the tool that cyborgs use to take actions, like submitting visit data to the state Medicaid system and sending a text message to a caregiver.
This is crucially important, because it keeps Mission Control as simple as possible. Simplicity scales.
Every Agent Makes Mistakes: Reinforcement Learning From Human Feedback
Cyborgs are half human at the end of the day. They also can make mistakes. All that matters is that they make fewer mistakes than humans.
In the great words of George Bush, "you fool me, you can't get fooled again."1The full clip is worth watching: www.youtube.com/watch?v=Hl7FKfl3O2Y We need to build a system where every mistake that an agent makes results in a recursive self improvement, where the agent doesn't make that same mistake again.
The cyborgs are finally here. Over the coming months, this company is going to change considerably. We'll shift from most of the work being done by humans, to most of the work being done by cyborgs. The future is here.