01
Understand
Classify intent and build a policy-aware plan.
TheOne
One agent on the surface. An intelligence operating system underneath.
TheOne is OneAI Labs' persistent root intelligence: a governed, verifiable agent kernel that understands human intent, acts across digital and physical systems, and learns from every verified outcome.
Not another chatbot
TheOne is not a demo shell. It is a kernel.
TheOne does not aim to be another chatbot, workflow builder or automation tool — those are capabilities it can use, not what it is. Its identity is a persistent root intelligence: it keeps its own state, memory, goals and responsibility chain across sessions, so closing the window does not reset what it was doing or what it has learned.
“A governed, verifiable and persistent intelligence that understands human intent, acts across digital and physical systems, learns from verified outcomes, and improves only through controlled release gates.”
The permanent core loop
Every capability serves the same cycle.
No action becomes trusted experience without verification, and no improvement reaches production without evidence.
02
Act
Route work to sub-agents, tools and connected systems under policy.
03
Verify
Score real outcomes against proof, cost and quality checks.
04
Learn
Turn a verified run into retrievable, reusable experience.
05
Evolve
Propose, evaluate and gate candidate improvements against a baseline.
Seven-layer architecture
Models and tools change. These seven layers hold identity and trust steady.
TheOne is built to let providers, models and product surfaces keep changing underneath it, without identity, governance, verification or accountability drifting.
01
Root Intelligence
Understanding, reasoning, decisions, delegation, reflection and adaptation.
02
Trusted Kernel
Identity, state, runs, policy, approval, budget, permission, scheduling and recovery.
03
Cognitive Plane
Model routing, planning, reasoning and capability self-awareness.
04
Memory & Experience
Context, episodic memory, semantic memory and an experience graph.
05
Agent & Execution Plane
Skills, sub-agents, durable runs, OneClaw, desktop, API, code and OneField.
06
Trust Plane
Verification, evaluation, artifacts, receipts, evidence, proof and audit.
07
Evolution Layer
Reflection, failure mining, candidate capabilities, benchmark, canary, promotion and rollback.
Where it stands today
The L4 foundation is wired — with explicit boundaries.
TheOne now connects durable execution, governed authority, memory, experience, a self model and independent verification. That is an L4 learning foundation, not a claim of autonomous self-evolution or unsupervised production promotion.
L1Assistant
Conversation and model calls.
L2Agent
Tool use and an agent loop.
L3Agent OS
Durable runtime, memory, policy and verification.
L4Learning Agent OS
Governed experience, self model, failure intelligence, adaptive routing and independent verification.
CurrentL5Self-Evolving Intelligence
Governed evolution of skills, strategy, prompts and model candidates.
L6Intelligence Organization
Dynamic multi-agent, long-horizon goals, enterprise-scale operation.
Verified today: the L4 foundation and its governed boundaries. Not claimed: autonomous self-evolution, unsupervised production promotion, or an independently proven L5 system. Repository evidence, tests, releases and production status remain the source of truth.
Current engineering focus
A governed learning foundation — not an autonomous one.
Every entry point — chat, automation, event, API, coding and multi-agent — resolves into the same governed run. Verified outcomes may become experience; candidate improvements still pass evaluation, approval, canary and rollback gates.
One Run. One State Machine. One Trust Model. One Recovery Model.
Strict run state machine — illegal transitions like completed → executing are rejected by default.
Immutable run events — the full, tamper-evident history behind every run, not a mutable log field.
Run-bound approval — a grant is checked against tenant, user, run, action and policy version, and cannot be reused by another task.
Full checkpoint contract — state, plan, tool state, budget, pending approvals and artifacts, so a run survives a restart or a deploy.
Unified entry points — automation, webhook, API and chat all create the same run and hand it to the same kernel.
Governed experience — only verified outcomes can become reusable experience, linked back to their evidence.
Controlled evolution — candidates are evaluated against a baseline, approved, canaried, monitored and reversible.
Build for production
Your AI should not stop at a demo.
Route intelligence through Core, govern it with TheOne, persist the work in OneMission, execute through OneClaw and keep verified outcomes in OneField.