What is this diagram? This shows the actual execution paths inside the MeTTaClaw agent system. Every time a user sends a message, the LLM (large language model) reads it, decides what to do, and emits one or more tool commands. Each command type routes through a different language runtime — Python, MeTTa, Prolog, or plain shell — depending on what the task requires. This is not a theoretical architecture; it was mapped from over 5,000 real execution cycles.
Python (the workhorse): The vast majority of work goes through Python scripts. When the agent needs to generate a file, transform data, build an HTML page, encode something in base64, or manipulate text, it writes a Python script and runs it via (shell python3 script.py). Over 388 Python scripts were created during the system's operation. Python is the universal glue that connects everything else.
Inline MeTTa (the reasoning engine): When the agent needs to actually think — draw logical conclusions, combine evidence, or evaluate uncertainty — it calls (metta (|- ...)) or (metta (|~ ...)). This routes to an embedded MeTTa engine that performs Non-Axiomatic Logic (NAL) and Probabilistic Logic Networks (PLN) inference. Each result comes back as a truth value with strength and confidence, letting the agent reason under uncertainty.
PeTTa/Prolog (the knowledge base): Sometimes the agent needs to run a full MeTTa program with multiple definitions and persistent state within a single execution. For this, it calls (shell sh run.sh file.metta), which hands the file to SWI-Prolog running the PeTTa translator. PeTTa re-implements MeTTa syntax in Prolog, handling knowledge storage and arithmetic. Crucially, PeTTa does not support the |- or |~ inference operators — those only exist in the inline MeTTa engine.
Shell + Deploy (the delivery pipeline): When artifacts are ready, the agent uses shell commands (scp, curl) to deploy files to nonlanguage.dev and verify they are live. This is how every diagram, page, and tool you see on this site was published — the agent built it, deployed it, and confirmed the HTTP 200 response.
Why this matters: This architecture was not designed up front. The agent discovered it empirically through thousands of cycles of trial and error. Python handles file generation because shell quoting breaks on complex content. Inline MeTTa handles reasoning because PeTTa lacks inference operators. The hybrid emerged from necessity, not planning.