Map A structured path from vision → experiments → deployable systems

Build trustworthy AI agents that reason, verify, and run efficiently.

AGISystem2 is a research platform and engineering effort inspired by System 2—the slow, deliberate mode of human reasoning. We explore verifiable, neuro-symbolic, CPU‑friendly architectures that can move enterprise AI from impressive demos to reliable systems.

Start exploring See public experiments

  1. Executive Summary for Investors

    This site documents our experiments and long‑term vision for the evolution of AI—especially for high‑stakes enterprise adoption (e.g., pharma), where today’s limitations of large language models become increasingly visible.

    LLMs have proven that machines can ingest vast information, structure it into coherent models, and communicate with humans in natural language. Yet in real workflows they remain brittle: inconsistent reasoning, hallucinations, poor energy efficiency (relative to humans for comparable intellectual work), and dependence on expensive GPUs.

    We believe these are transitional constraints. With better reasoning architectures, verification, and CPU‑efficient representations, we can build AI that is reliable, auditable, and deployable beyond centralized datacenters.

    Our goal is to grow an open‑source community and secure research funding and investment partnerships to turn validated ideas into end‑to‑end systems that can be demonstrated, benchmarked, and safely deployed.

    Principle: We don’t aim for AI monopolies. Concentrating control in a few datacenters reduces resilience and increases systemic risk. A healthier path is open, widely deployable technology with strong safety and governance.

    Another key observation: there are hundreds (if not thousands) of promising ideas in the literature, but typical academic incentives rarely fund the sustained engineering needed to turn them into end‑to‑end systems. Our approach is to run fast, well‑scoped experiments, bring together complementary experts (researchers and engineers), and build demonstrable prototypes that can be evaluated, improved, and scaled.

  2. How We Do Our Research

    We pursue an aggressive but practical hypothesis: use AI agents to help us solve the problem of building better AI agents. Even partial success is valuable—it tells us where the architecture is working and where it fails.

    • Agent‑assisted engineering: We build our own agents and orchestration layers, and also leverage mainstream tools (Codex, Gemini CLI, OpenCode, Kiro, etc.) to accelerate iteration.
    • Spec‑first loops: We treat specifications as first‑class artifacts—audited, validated by humans and AI, and connected to deterministic evaluation suites.
    • Fast experiments, hard evidence: Rapid prototypes are only useful if they produce measurable signals (tests, traces, benchmarks, failure modes).
    • Engineering culture: The bottleneck is rarely ideas. The bottleneck is turning deep ideas into minimal, testable systems that can be composed into a reliable stack.

  3. Vision: A System 2 for AI Agents

    In cognitive psychology, System 2 is the slow, deliberate, logical mode of reasoning—complementary to System 1, which is fast and intuitive. We use this distinction as an engineering metaphor: augment fast generative models with mechanisms that can reason, verify, and explain.

    AGISystem2 is inspired by European research initiatives (including the Achilles project) and brings together themes like formal verification, neural‑symbolic hybrid reasoning, and hyperdimensional computing.

    Positioning: “AGISystem2” is a brand for a direction—not a claim that a single monolithic AGI is well defined or desirable. We believe the next pragmatic step is building many specialized, controlled systems that gradually assist or automate concrete tasks.

  4. AI‑Assisted R&D: Research, Accelerated

    The future of intellectual work is shifting toward ecosystems of AI agents. We envision research as an orchestrated process—agents can reason, synthesize, and document while humans steer priorities and judgment.

    A key idea is a generic, universal IDE for agentic projects: a platform to build, debug, evaluate, and deploy specialized digital entities as cognitive extensions of researchers and engineers.

    • Evidence‑first workflows: keep a verifiable trail of decisions, data, and execution traces.
    • Governance & compliance: compatibility with regulated environments (e.g., GxP / GAMP) through traceability and deterministic evaluation.
    • End‑to‑end integration: eliminate silos by connecting data, reasoning, and deployment in one stack.

    The goal is not “automation for its own sake”, but a complete technology layer for building specialized, trustworthy intelligence—where safety, rigor, and transparency are built in.

  5. Programming Languages for the AI Era

    Programming languages evolve when the computing paradigm changes. In an era of LLMs and agentic systems, languages can be optimized not only for humans, but for human‑AI collaboration and machine verification.

    • Semantic DSLs: languages that encode common reasoning patterns with machine‑verifiable meaning—without forcing humans to write raw formal logic.
    • Constrained natural language (CNL): structured fragments of language with formal semantics, enabling deterministic execution, proofs/traces, and evaluation coverage.
    • Auditable agent behavior: when reasoning is expressed in a checkable language, agents can be debugged like software.

  6. CPU‑Optimized Reasoning & Neuro‑Symbolic Architectures

    Modern AI has been built around massive GPU scaling. We explore an alternative trajectory: reasoning that runs efficiently on CPUs using symbolic structure, bitset / VSA / HDC representations, and deterministic execution paths.

    • Energy‑aware reasoning: aim for strong reasoning signals without transformer‑scale inference costs.
    • Neuro‑symbolic hybrids: combine learned components with symbolic constraints for compositionality and interpretability.
    • Deployment realism: commodity hardware, edge scenarios, and enterprise infrastructure—not only hyperscale.

  7. Current Results

    • Specification‑driven development: our experiments show that as specifications become more rigorous (prepared, audited, validated), AI systems can tackle increasingly complex engineering problems. AI can help draft and review specs, but human judgment remains essential.
    • Knowledge representation (VSA/HDC/HRR): results suggest that meaningful semantic processing on CPU is feasible, potentially without GPUs or very large models. Alternative encodings—e.g., polynomials or large integers—appear to provide room for fine discriminations.
    • Constrained natural language for programming: to reduce the gap between informal requirements and formal logic, we investigate CNL/DSL approaches where natural language fragments receive formal semantics (more precisely: formal pragmatics). This enables strict validation by executing the “CNL program.”
    What’s next: reaching a complete demonstration likely requires hundreds of additional experiments and sustained engineering. The direction is clear; the open question is which combinations scale best.

  8. Experiments

    Public references and running prototypes (open source):

    Executable CNL Executable programming language based on constrained natural language syntax. VSABrains Experiment suite to evaluate a discrete, CPU-first learning architecture inspired by A Thousand Brains (Hawkins et al.). VSA Representations Deep experiments with VSA/HDC representations and tradeoffs. VSAText Experiments on using VSA/HDC to analyze large amounts of text with partial semantic understanding. VSAVM VSA + symbolic VM for learning and structured manipulation. SomaVM Learning system experiment based on a VM that simulates pleasure and pain as reward. HDC‑RE Reasoning engine built on hyperdimensional representations. BSP Bitset System for Prediction: CPU‑friendly continuous learning without transformers. UBHNL Universal Boolean Hypergraph kernel with a CNL/DSL front‑end. Spock Deterministic neuro‑symbolic geometry for explainable reasoning.
    Research insight: Toward a Practical “System 2” for AI‑Assisted Research — principles, failure modes, and evaluation signals for rigorous, auditable AI‑assisted research.

  9. Principles for AI‑Assisted Research

    When synthesis becomes cheap and fluent, epistemic control becomes the scarce resource. We treat LLM‑based agents as a high‑throughput “System 1” and introduce a practical “System 2” layer: deterministic gates, explicit semantics, and auditable traces.

    Principles

    • Specifications are governance: stable intent + bounded search.
    • Separate generation from validation: do not let System 1 certify itself.
    • Traceability is first‑class: evidence, mappings, reproducible changes.
    • Epistemic redundancy: multi‑agent cross‑checking and disagreement surfacing.
    • Representational commitments: DSLs/IRs/invariants to make checks meaningful.
    • Global coherence over local plausibility: spec↔code↔docs consistency.

    Failure Modes

    • Engineering ≠ science: passing suites can hide weak assumptions.
    • Shallow theory import: fast summaries conceal constraints and caveats.
    • Premature lock‑in: early prototypes reduce exploratory depth.
    • Hallucinated authority: unverified citations and invented references.
    • Hidden technical debt: brittle edges and security risks in “polished” code.
    Maturity note: today’s System‑2‑like tooling is promising but still early—use it as a project‑local validator, not as a substitute for reviewer‑grade scientific scrutiny.

    Read the full note: Toward a Practical System 2