Ross Eyre

Co-founder @ArkologyStudio

[Ross Eyre @rosseyre]

@forwarddeploy Currently building HabitatsOS, cooperative environments for humans + agents. Token budgets are the biggest constraint for our team rn!

[Ross Eyre @rosseyre]

https://t.co/xghtFVZTYf

[Ross Eyre @rosseyre]

@MattPRD @MarikHazan Hi Matt – see your dm’s r.e multi-agent coordination infra. Think you’ll like this one.

[Ross Eyre @rosseyre]

The representational substrate for machine participation. Incredible essay by @che_coelho – with both philosophical and technical depth

[Arkology Studio @ArkologyStudio]

New post ✉ 'Designing Coherent Interfaces Between Cybernetic Systems and Socio-Technical Organisms'

In it @che_coelho sketches a blueprint for coherent cybernetic interfaces to socio-technical organisms—optimising for coherence over omniscience.

https://t.co/q2MVTZpfel

[Ross Eyre @rosseyre]

@RubenLaukkonen Can’t wait to dive into this. Allocated to tonight’s reading 📚

[Ross Eyre @rosseyre]

What does this mean for software?

In part it's a shift in power. Skills essentially offer a way for non-engineers, specifically domain experts, to ‘program’ agentic systems in natural language.

This isn't ‘vibe coding’ or ‘prompt engineering’, it’s about using language to describe a constraint space within which a reasoning agent operates toward a target solution.

If you’re reading this and have deep domain expertise, consider composing skill packs that encode the core activities of your workflows. You may have just described a set of behavioural attractors for a future vertical AI SaaS agent.

[Ross Eyre @rosseyre]

Most people still think the power of AI agents comes from the model. But the model is just the reasoning substrate.

How that reasoning is orchestrated toward an end goal is what differentiates agents that drift and those that finish tasks with high competency. 🧵

[Ross Eyre @rosseyre]

By providing a set of behavioural attractors, you guide the agent toward the target while allowing it enough freedom to improvise within the constraints.

In my last post, I likened this quality to genetic instructions: compact programs that reliably shape behaviour but which don’t over-specify the result or expression.

[Ross Eyre @rosseyre]

Interesting thesis on general-purpose v.s. vertical AI agents.

The thesis (in my own words):
Program synthesis (i.e the process of discovering how to behave by interacting with the world and adjusting based on what happens) is a general problem-solving engine.

Once an agent can:
•represent goals
•search programs
•execute them
•learn from outcomes

…it no longer needs:
•bespoke ontologies
•handcrafted workflows
•domain-specific intelligence

This tracks how we’ve been thinking about agent design at @ArkologyStudio.

An agent that can manipulate its ‘habitat’ (i.e computational environment) by writing and executing code, using scratchpad artefacts to offload memory/cognition, and call tools and APIs – all within a self-validating loop – is a much more capable and adaptive system for doing just about any ‘knowledge’ work.

Food for thought for anyone working on AI agents in 2026 ..

[Aakash Gupta @aakashgupta]

Amjad is onto something most AI labs haven’t fully internalized yet.

I’ve spent the last year watching verticalized agent startups raise hundreds of millions only to get steamrolled by generic coding agents.

The pattern keeps repeating.

Legal AI agent? Claude Code can read statutes and draft contracts. Medical documentation? Coding agent with the right prompts handles it. Financial modeling? Same story.

The verticalized approach assumes domain expertise is the moat. But Sutton’s bitter lesson says the opposite. General methods that scale with compute always win.

What makes coding agents different is they have somthing no verticalized agent has. A universal interface to the world.

Code can call any API. Code can parse any document. Code can automate any workflow.

Every vertical agent is basically a worse version of a coding agent with guardrails.

The VC math here is BRUTAL.

Enterprise AI startups raised something like $8B in 2024 building bespoke solutions for healthcare claims processing, insurance underwriting, legal discovery.

Meanwhile Cursor hit $1B ARR in 36 months just letting developers write code faster.

And here’s where Amjad’s insight gets really interesting.

He’s right that program synthesis maps exactly to the bitter lesson. Search and learning. Those are Sutton’s two scalable methods.

What is a coding agent doing?

Searching the solution space of all possible programs. Learning from feedback loops. Executing and iterating.

Every handcrafted vertical agent is the equivalent of 1990s chess programs encoding grandmaster knowledge. They work until they don’t.

AlphaZero didn’t need chess theory. Claude Code doesn’t need to be trained on legal precedent.

Microsoft is catching up to this realization late. They built an empire of verticalized Copilots for Sales, Service, Finance.

Replit figured it out early. Give developers a general coding environment and let them build whatever they need.

The winners in AI infrastructure will be the ones who understood program synthesis was the endgame all along.

[Ross Eyre @rosseyre]

TTT-E2E is a step toward continual adaptation in AI systems. Biology solved this eons ago. It had to!

[Ross Eyre @rosseyre]

What if AI models could learn while they're talking to you?

New paper from Stanford University & NVIDIA demonstrates a novel continuous learning system for LLMs, blurring the lines between ‘training’ and ‘inference’.

Make sure you're sitting down for this one 🧵

[Ross Eyre @rosseyre]

Here’s a mental model: Instead of thinking about "training vs. inference", as separate phases, think global vs. local learning

Global: Slow, pre-training phase. Builds general capabilities, shared across users/tasks.

Local: Fast, during inference. Tailored to this doc/user/task—ephemeral, situational.

TTT-E2E ultimately enables local learning during inference, making LLMs more context-aware and personal.