What research areas does XAGI Labs work in?

Autonomous AI agents, long-term memory systems, agent swarms, computer-use agents, self-evolving tool synthesis, world models, active inference, and the broader path to accessible AGI.

Does XAGI Labs publish papers?

We publish research notes and engineering deep-dives on the Research page as our systems mature. We prioritize shipped systems over preprints — every claim is grounded in production ATLAS code.

Research · Lab notes

Notes from the workbench.

Working notes — not papers — on what we've learned shipping ATLAS. Honest, dated, and occasionally wrong. We publish when the lesson is durable, not when it's clever.

Lab notes

What we're learning.

2026 · May

Why Rust isn't a free lunch for small-input hot paths.

Native / FFI

2026 · Apr

Sleep gates, surprise gates, and what not to remember.

Memory

2026 · Apr

Closed-loop verifier: tool synthesis that actually retains.

Tool synthesis

2026 · Mar

Spec-mode, plan-mode, act-mode: state machines for autonomy.

Agents

2026 · Mar

Semantic tool routing with a 270M model. Yes, really.

Routing

2026 · Feb

Kernel-sandboxed tool execution across three operating systems.

Safety

2026 · Feb

Why the desktop, not the chat window, is the right surface.

Product

2026 · Jan

Founding note: a continuously-running mind, not a chatbot.

Vision

Notes are pre-print drafts. Reach out at research@xagilab.com to discuss any of them.

Areas

What we're working on.

Long-horizon autonomy.

Agents that survive across hours, days, and the messy edges of real work — picking up where they left off when things go wrong, instead of starting over.

Living memory.

What does it mean to remember well? We're studying the shape of memory that helps a system stay useful for years — and what to let go of.

Skills it can grow.

When ATLAS encounters something it can't yet do, it teaches itself — safely. Quietly. We're working on making that loop quick, trustworthy, and personal.

Eyes on the screen.

Letting ATLAS see what's on your screen and act with care — accurate, reliable, and graceful even on older hardware. Cross-platform from day one.

Quiet speed.

Performance that gets out of the way. The hard part isn't making things fast — it's knowing exactly where speed actually matters to a person using the product.