Contents

• The Problem
• MeTTaClaw: The Foundation
• HyperClaw: Three Additions
• Two Versions, One Architecture
• The Scoring Logic
• Decentralized Deployment on ASI:Chain
• Why This Matters

Ben Goertzel has published a new preliminary design proposal titled HyperClaw: Cognitive Orchestration via Attention-Metaprotocol for Hybrid AGI Systems, and if you are serious about where AGI architecture is heading, it deserves your full attention. 

[Read the full paper here]

Dr. Ben Goertzel recently shared some early thinking on HyperClaw in the video discussion below. If you work in AI research, agent systems, or are following the road to AGI, this is a design proposal worth reading carefully. We’ve broken down the key ideas below to give you a clear sense of what it contains and why it matters.

The Problem

Anyone who has worked seriously with multiple AI tools at once knows the frustrating reality: the human is the glue. You decide which model to call, what context to pass, when to pivot strategy, and how to reconcile contradictory outputs from different systems. The rote coordination is time-consuming but uncreative. The genuine strategic judgment (knowing when your whole approach is wrong and you need to try something fundamentally different) takes only a few hours but is irreplaceable.

The paper illustrates this with a quantitative finance research workflow involving roughly 100 experimental iterations across notebooks, papers, and data pipelines, consuming around 30 hours of human time. Most of that is mechanical: generate/critique/revise cycles, bug fixes, parameter sweeps. Only a few hours involve real creative intervention: recognizing the need to pivot from volatility-based regime definitions to compression-based ones, or catching the moment when an LLM quietly substitutes a hallucinated method for the one that was actually validated.

HyperClaw is designed to automate the mechanical majority, leaving the researcher with the strategic minority.

MeTTaClaw: The Foundation

HyperClaw is built on top of MeTTaClaw, a compact agentic system implemented entirely in MeTTa (SingularityNET’s AGI programming language) by Hyperon team member Patrick Hammer. The entire agent core is about 200 lines of MeTTa code, making the control loop fully inspectable and modifiable.

MeTTaClaw provides three core capabilities: embedding-based long-term memory stored in MeTTa’s Atomspace format, tool use (web search, file access, shell commands), and skill acquisition through memory extensions. It is already operational and deployed as a lightweight agent.

Two properties make MeTTaClaw more than just a tidy reimplementation of OpenClaw. First, MeTTa is designed for self-rewriting code, meaning a system built in MeTTa is, in principle, a system that can modify and improve its own architecture. Second, MeTTaClaw already inhabits the same Atomspace substrate that Hyperon’s full AGI cognitive modules will use. This means any orchestration layer built on top of it can grow from LLM coordination today to AGI cognitive process coordination tomorrow, without a change of substrate.

HyperClaw: Three Additions

HyperClaw extends MeTTaClaw with three architectural elements.

Context Frames are typed MeTTa structures representing the complete state of a multi-module task: active goals, working hypotheses, experimental history, module performance ratings, evaluation criteria, and, critically, a certified record of methods that have actually been validated through experiment. Every module invocation receives a relevant projection of the current context frame and writes its output back as a frame update. The context frame, not LLM memory, is the authoritative source of truth. When the system builds a final deliverable, it must reference the certified method entry in the frame. It cannot substitute what an LLM happens to recall, which may be a hallucination.

The Attention-Metaprotocol is a set of MeTTa rewrite rules governing which modules get attention, when, and with what context. It runs at two timescales. The fast loop (seconds to minutes) automates rote orchestration: generate/critique/revise cycles, bug-fix loops, parameter sweeps, code execution. The slow loop (minutes to hours) handles strategic steering: evaluating whether the current direction is working, proposing mode switches between strategic postures like “explore methods,” “refine parameters,” “validate,” and “wrap up deliverables,” and flagging pivots that require human approval. The protocol is “meta” not because it tells modules what to compute, but because it governs what they attend to and how their outputs are admitted into the shared context.

Module Spaces wrap every external system (LLM endpoints, code executors, data stores, sensor streams) behind a uniform Hyperon interface. The orchestrator interacts with all of them identically, regardless of whether a module is a commercial API, a locally hosted open-source model, a Python process, or eventually a full PRIMUS cognitive subsystem. Adding a new module is just adding a new Space.

Two Versions, One Architecture

The paper’s most intellectually interesting argument is that the same architectural skeleton can serve two quite different purposes.

HyperClaw v1 is a near-term practical system. It orchestrates multiple LLMs and code executors for complex research workflows, with a human providing high-level strategic direction. The paper details two concrete applications: the quantitative finance workflow described above, and a humanoid classroom robot that must simultaneously manage perception, dialogue, motor control, emotional appraisal, and safety monitoring across multiple AI backends in real time. For the finance workflow, the projected reduction is from ~30 hours of human effort to ~5 focused hours of strategic input.

HyperClaw v2 deepens the same architecture into a full AGI cognitive orchestration layer. The context frames, attention-metaprotocol, and module Spaces stay exactly the same. What changes is what sits behind the module Spaces. Instead of LLMs and code executors, v2 orchestrates the full PRIMUS toolkit: PLN for probabilistic reasoning, MOSES for evolutionary program synthesis, ECAN for attention allocation, WILLIAM for pattern mining, predictive-coding neural networks, MetaMo for motive geometry, and SubRep for subgoal management. The creative strategic interventions that v1 hands off to a human are handled in v2 by WILLIAM’s pattern mining and concept blending.

The upgrade path is explicit and incremental. Individual components can be swapped from LLM-based to PRIMUS-based one at a time, with the system falling back to LLM processing for any component where the PRIMUS replacement isn’t mature yet. HyperClaw v2 is not a different system; it is v1 with deeper cognitive machinery behind the same interface.

The Scoring Logic

The fast loop’s step selection is worth understanding concretely. Each candidate action is scored by the product of two factors: forward reachability (how feasible is this step given current resources and module capabilities?) and backward usefulness (how much does this step advance the active goals?). The system selects the action maximizing the gain in the sum of the logs of these two factors per unit of cost, which the paper calls geodesic scoring.

This isn’t just a technical detail. The same geodesic control principle governs PLN inference, MOSES evolution, and knowledge transfer throughout PRIMUS. By using it for orchestration decisions as well, the architecture ensures that the system’s coordination logic is mathematically consistent with all of its other cognitive decisions, making behavior more interpretable and auditable across the board.

Decentralized Deployment on ASI:Chain

The paper also lays out how HyperClaw is designed for eventual deployment on ASI:Chain, SingularityNET’s forthcoming decentralized AI infrastructure. In that context, context frames and certified method records become content-addressed objects with cryptographic provenance, meaning the question “what method did the system actually use?” always has a verifiable answer. Module Spaces run under capability contracts enforced by the Rholang runtime, meaning a module literally cannot access frame data outside its permitted scope. New orchestration strategies deploy first in shadow mode, run in parallel for output comparison, and graduate to primary use only when KPIs stabilize.

The multi-party implications are significant: different teams can contribute module Spaces under formal capability guarantees, DAOs can gate upgrades to shared orchestration rules based on published performance certificates, and the attention-metaprotocol’s geodesic scoring naturally creates economic pressure toward efficient orchestration since compute and API costs are metered transparently.

If you’re looking for a place to build in the next phase of the stack wars, the ASI Chain DevNet is it. [Review documentation] [Visit website]

Why This Matters

What makes HyperClaw worth paying attention to isn’t any single component; it’s the architectural continuity. Most AI orchestration frameworks today are designed around LLMs and have no natural path toward richer cognitive architectures. HyperClaw is explicitly designed so that the same interface that today wraps a Claude or GPT-4 endpoint can tomorrow wrap a PLN inference engine or a MOSES evolutionary run. The system doesn’t need to be rebuilt as the underlying cognitive modules mature; it just needs new Spaces.

The paper is explicitly a preliminary design proposal marked for discussion, not a finished system. But as a blueprint for how you might actually build a cognitive orchestration layer that scales from practical automation today to integrated AGI tomorrow, it’s a serious and detailed piece of thinking.

HyperClaw is under active development at SingularityNET / ASI Alliance / Hyperon. The design proposal is preliminary and subject to revision.