Neuro-symbolic agent

An agent loop in which the LLM does the language work (reading messages, turning text into formal statements, choosing which rule to apply and when to stop) and a separate formal engine does the inference, with truth values the model can set but not recompute. The word for the pattern predates LLM agents by decades; in this spoke it names a design point on the same axis as agent-skills (procedure as markdown the model follows) and jaseci's meaning-typed programming (the type system as the prompt): what part of the agent's reasoning is taken out of the model's hands.

The one instance, and what it found

omega-framework (singularitynet, 2026) is the corpus's first. Its call-and-wait cycle suspends generation whenever the model emits (metta (|- ...)), runs Non-Axiomatic Logic or Probabilistic Logic Networks over an AtomSpace, and injects the result atoms, (robin --> animal) (stv 0.72 0.32), back into the next prompt as data. The claim is auditability: "When the agent says it is 72% confident, that number comes from formal inference — not a feeling."

The project's own failure catalogue is the useful part, because it locates where the design moves the failure instead of removing it:

• Upstream of the engine. The model writes the premises, and on the reference deployment LLM-supplied facts were right ~55% of the time, relations were inverted up to 16.6% of the time, and self-assigned confidence ran ~15 points high. The engine cannot see any of this.
• Through the engine. Deduction multiplies confidences with no floor, so a chain is below the action threshold by hop three, and a wrong premise comes out with a rigorous-looking truth value; the doc's phrase is "GIGO amplification."
• At the interface. The single most common failure over 4,500 cycles was the model not producing a parseable s-expression at all.

So the formal layer buys one thing exactly, that the model cannot inflate a derived confidence by rhetoric, and pays for it with a new failure surface at the boundary, plus the same garbage-in problem every other harness has, now wearing a number. Whether that trade is worth it is not measured anywhere yet: the only evaluation is the framework's own agent measuring itself.

Open questions

• Is there an independent comparison of a neuro-symbolic loop against a plain LLM loop on the same tasks? Omega's numbers are Omega-on-Omega.
• The action thresholds (ACT at f ≥ 0.6 ∧ c ≥ 0.5) gate on the truth value of a conclusion rather than on the kind of action, which is what every agent-permission-model in this spoke gates on. Are the two dials complementary or does one subsume the other?

Related

omega-framework · agent-framework · agent-skills · jaseci · agent-permission-model · self-improving-agents · agent-memory