// LABORATORY PROJECT
An agent that does not need to be told what to do.
Argos is OrvixLabs' laboratory project. Most current agents are executors: they receive a mission defined by a human and fulfill it with greater or lesser sophistication. Argos inverts the problem. Argos doesn't need to be told what to do: it observes its environment, builds a model of it, sets its own goals from that observation, and plans with continuous self-correction. Its mission is not an input: it is a conclusion it reaches.
The capacity to generate mission, not just to fulfill it.
// OrvixLabs is a hybrid agency. Argentine systems studio and technical consultancy based in Necochea, Buenos Aires. We design, build and operate private infrastructure with enterprise artificial intelligence.
// AGENTS THAT DON'T WAIT FOR INSTRUCTIONS
When the agent decides what to do.
Today's agents are executors. They get a mission, fulfill it, wait for the next one. Argos inverts the problem. Observes its environment, models the world, derives its own objectives and plans with self-correction. No commercial deadline. OrvixLabs' bet on what's coming.
// THE FOUR CAPACITIES
What separates Argos from any existing framework.
Argos is defined by the simultaneous coexistence of four properties. Any system demonstrating one or two can be called an assistant; only when the four operate in a loop does something emerge that deserves the name of an autonomous reasoning agent.
Autonomous environment exploration
Argos actively probes its context instead of waiting for instructions. It identifies what is there, what changes, what patterns repeat and what signals are new. Exploration is a constant process, not an initial phase.
World modeling from observations
The observed condenses into an internal representation: entities, relationships, dynamics, regularities. The model is not static; it is revised every time reality contradicts the prediction. The model is the raw material for every subsequent decision.
Goal setting without explicit instructions
From the model, Argos derives what is worth doing. It detects gaps between the current state and desirable states, prioritizes, and proposes its own goals. Here is the qualitative leap: the mission emerges from reasoning, not from the prompt.
Planning with real-time self-correction
Once a goal is set, Argos builds a plan, executes it and rewrites it live when environmental feedback doesn't match expectations. Correction is not an exception but the default operating mode.
The four capacities must operate as a closed loop: explore → model → set goal → plan → correct → explore again. If any breaks, Argos collapses into a conventional assistant.
// GUIDING PRINCIPLE
Margaret Hamilton: designing assuming everything can fail.
Hamilton designed the Apollo 11 software starting from the premise that any subsystem could collapse at any moment, and built a priority architecture capable of continuing to operate when that happened. Argos inherits that philosophy: every layer, sensors, model, goals, plans, actuators, is designed with the hypothesis that the others can fail, return garbage, or become outdated.
In practical terms, this defines concrete architectural decisions: graceful degradation, cross-verification between components, state persistence, and a default behavior that is always safe when the system enters into doubt.
// FIRST APPLICATION
Smart home central agent.
The first deployment environment is the home. Not because it is trivial, it isn't, but because it combines four ideal conditions to validate the capacities: a bounded environment, rich in heterogeneous signals, with a human user providing continuous feedback, and with real consequences if action is wrong.
Natural voice
Conversational interaction as primary interface, not as command.
IoT management
Domestic hardware control: lighting, climate, security, appliances.
Cognitive assistance
User support in planning, memory, everyday decisions.
Anticipation
Need prediction from observed patterns, not from hardcoded rules.
The home is the first domain, not the last. Argos is designed environment-agnostic: the same architecture must be transferable to an industrial environment, a vehicle, or a digital context without conceptual rewrite.
// THEORETICAL FRAMEWORK
Argos is not science fiction.
Argos' four capacities rest on mature bodies of artificial intelligence research. This is not an exhaustive list but an entry guide for those who want to go deeper into each pillar.
Autonomous agents and cognitive architectures
- Russell, S. & Norvig, P. (2020). Artificial Intelligence: A Modern Approach, 4th edition. Pearson. Chapters 2 (Intelligent Agents) and 26 (Robotics).
- Wooldridge, M. (2009). An Introduction to MultiAgent Systems, 2nd edition. Wiley.
- Sumers, T., Yao, S., Narasimhan, K. & Griffiths, T. (2024). "Cognitive Architectures for Language Agents". Transactions on Machine Learning Research.
World models and internal representations
- Ha, D. & Schmidhuber, J. (2018). "World Models". arXiv:1803.10122. Foundational work on agent-learned world models.
- Hafner, D., et al. (2023). "Mastering Diverse Domains through World Models" (DreamerV3). arXiv:2301.04104.
- LeCun, Y. (2022). "A Path Towards Autonomous Machine Intelligence". Open Review. JEPA architecture and predictive world model agents.
Intrinsic motivation and autonomous exploration
- Pathak, D., et al. (2017). "Curiosity-Driven Exploration by Self-Supervised Prediction". ICML 2017.
- Schmidhuber, J. (2010). "Formal Theory of Creativity, Fun, and Intrinsic Motivation". IEEE Transactions on Autonomous Mental Development.
- Oudeyer, P-Y. & Kaplan, F. (2007). "What is Intrinsic Motivation? A Typology of Computational Approaches". Frontiers in Neurorobotics.
Autonomous goal-setting
- Colas, C., et al. (2022). "Autotelic Agents with Intrinsically Motivated Goal-Conditioned Reinforcement Learning". Journal of Artificial Intelligence Research.
- Stooke, A., et al. (2022). "Open-Ended Learning Leads to Generally Capable Agents". DeepMind. arXiv:2107.12808.
- Wang, R., et al. (2019). "POET: Open-Ended Coevolution of Environments and their Optimized Solutions". GECCO 2019.
Adaptive planning and self-correction
- Camacho, E.F. & Bordons, C. (2007). Model Predictive Control, 2nd edition. Springer.
- Sutton, R.S. & Barto, A.G. (2018). Reinforcement Learning: An Introduction, 2nd edition. MIT Press. Chapters on planning and planning-learning architectures.
- Yao, S., et al. (2023). "ReAct: Synergizing Reasoning and Acting in Language Models". ICLR 2023.
Margaret Hamilton and the Apollo 11 software
- Hamilton, M. (1976). "Hamilton's Lessons Learned: Apollo Software Development". NASA Internal Document.
- Mindell, D. (2008). Digital Apollo: Human and Machine in Spaceflight. MIT Press.
- Eyles, D. (2018). Sunburst and Luminary: An Apollo Memoir. Fort Point Press. Memoirs of the lunar module software co-designer.
// CURRENT STATUS
Where the project stands.
Project formally defined and named. Thesis closed. The four fundamental capacities are conceptually delimited. What remains is materializing the base architecture integrating them into a single operational loop. Argos is a laboratory project without commercial deadline, its priority is architectural solidity, not speed to market.
Build an agent that doesn't need to be told what to do, because it already understood, by itself, what needs to be done.
// FAQ
Frequently asked questions about Argos
What is Argos? +
Argos is the OrvixLabs laboratory project on autonomous reasoning agents. Most current agents are executors that receive a mission and fulfill it. Argos inverts the problem, observes its environment, builds a model of the world, derives its own objectives and plans with continuous self-correction.
Is Argos for sale? +
No. Argos is laboratory research without a commercial deadline. The priority is the architectural soundness of the four integrated capabilities (observe, model, derive objectives, plan with self-correction), not time to market. When an Argos product is ready for a company, we will announce it specifically.
What are the four capabilities of Argos? +
First, autonomous exploration of the environment without waiting for instructions. Second, world modeling from observations. Third, establishment of own objectives without explicit instructions. Fourth, planning with real-time self-correction when feedback from the environment doesn't match expectations. The four must operate in a continuous loop, not separately.
Why start testing it in a house? +
A house combines four ideal conditions, a bounded environment with heterogeneous sensory signals, recoverable errors without serious consequences, a human user providing continuous feedback, and enough complexity for the problem to be real. In an industrial plant, the consequences of an error are too large to start there. In a house, the agent learns fast and no one gets hurt.
When will Argos scale to industrial or vehicular environments? +
The conservative roadmap targets a functional home version by end of 2026 or early 2027, industrial during 2027, vehicular during 2028. Each leap is an order of magnitude technical step and timelines may shift based on architectural validation. We don't commit dates so the product is ready, we do it the other way around.
Does Argos share architecture with IRIS SCE or Casandra? +
Argos inherits principles from both but is a distinct architecture. It inherits the Hamilton philosophy from IRIS, assuming any subsystem can fail. It inherits Casandra's adversarial deliberation to validate the world-model conclusions. But the exploration-model-objectives-plan loop is proprietary to Argos.
What is the difference between an Argos agent and a Claude or GPT agent? +
Current LLM-based agents execute given tasks. You tell them what to do, they do it. Argos decides what to do. It doesn't wait for a mission. It observes, models, derives its own mission from observation, and fulfills it with self-correction. The difference isn't technical capability, it's cognitive initiative.
Does Argos comply with the Magnifica Humanitas principle? +
Yes, deliberately. Argos has initiative to act but critical or high-impact decisions are escalated to humans for validation. The autonomy is not total, it is operational within a framework where ultimate responsibility remains human.
// WHAT MOTIVATES US MOST
Argos in moments when humans cannot get there in time.
The application of Argos that motivates us most is not the domestic one. It is response to situations where fast, autonomous, environment-aware decisions can save lives. An agent that sees, models, infers objectives and acts without waiting for instructions has a natural place there, where sending a human is too slow, too dangerous, or simply impossible.
Natural disasters
Earthquakes, floods, wildfires, tsunamis. Argos operates where networks are down, where there is no continuous connectivity, where the situation changes hour by hour. It observes with its own sensors, models the affected area, prioritizes critical zones, and coordinates with whoever can actually get there.
Building emergencies
Building fires, collapses, evacuations. Argos can operate inside the building itself, read fire progress or structural damage, identify where people are trapped, suggest safe evacuation routes that change as the situation changes. Live information for the rescuers.
Conflict and humanitarian crisis
War zones, areas of active conflict, situations where human presence is extremely high risk. An autonomous agent can enter to observe, identify civilians, critical infrastructure, exit routes, without exposing operators. The difference from a traditional drone is that it decides on its own what to look at and what to report.
Operation in unviable environments
Nuclear plants with incidents, mines with collapse risk, compromised critical infrastructure. Where a human cannot enter or the cost of entering is disproportionate, an agent that decides what to do based on what it sees and learns from the environment changes the rules of the game.
// THE PRINCIPLE THAT MAKES IT POSSIBLE
Not a drone with AI. An agent with criterion.
A drone follows orders, goes where it is told, does what it is programmed to do. Argos observes, infers what matters in the context it faces, decides what to report and what action to prioritize, and self-corrects when what it predicted does not match what it sees. The difference is of category, not of power.
This sits within the Magnifica Humanitas principle. Argos deliberates, models, recommends and executes operational tasks. Decisions that affect lives always escalate to a responsible human. Autonomy is of movement and observation, not of ultimate judgment.