Most agents let the model drive every step: reading raw data, deciding, calling a tool, reading more. It’s expensive and non-deterministic. OpenSymbolic narrows the model’s job to writing a plan, then runs that plan as real code. Here’s exactly how that works.
In a standard agent loop, the language model is in the driver’s seat the whole way through, so it re-reads the same documents every turn, improvises each step, and gives you a different answer to the same question. OpenSymbolic inverts that. The model writes a plan once, in code, against a fixed set of typed building blocks. Ordinary Python runs the plan (retrieval, filtering, extraction, calculation), and the model only ever sees the small, relevant result at the end. The heavy lifting is deterministic code, so it’s cheap, repeatable, and traceable.
This is why it’s so much cheaper. During planning, the model works with variable names and short summaries, never raw page content. Documents stay in Python memory and move between steps as variables, whether that’s 100 rows or 100,000. Context grows ~0.5–2K tokens per step instead of the 5–30K a raw-data loop re-sends every iteration. In independent testing, this produced 86% fewer tokens with zero errors.
All three share the same primitives, the same security model, and the same plan-then-execute core. Swap patterns without rewriting your agent.
For deductive and constraint problems, the model translates the question into formal logic and a theorem prover (Z3) checks it. The model handles language; the prover handles truth. On the FOLIO first-order-logic benchmark this reached 89.2% (near the 96% human bar) with a 100% plan-success rate, because a prover catches what a model would otherwise guess at.
Because the model plans over variable names and can only call approved functions, whole classes of risk are closed off by construction rather than by prompt.
Prompt engineering plateaus: past ~70% accuracy, every fix introduces a new regression. OpenSymbolic replaces prompt archaeology with a structural loop. Improvements are additive and measurable, and they don’t break what already worked. This is why 7 of 11 models hit 100% on TravelPlanner: the framework handles the hard part.
The framework is open source: MIT, pip install opensymbolicai-core. Build a working agent in five minutes, read the generated plan, inspect the trace, see the token count. Swapping models is a one-line config change.