For the engineer

Send logic, not data.

If you were sent this page, someone wants you to check whether this is real. Here is the mechanism, the code, the guarantees, and the places to try it yourself.

The shift in control

Placeholder. A standard tool-calling loop lets the model decide every step and re-reads the full transcript each time. OpenSymbolic inverts this: the model plans once, code executes the plan, and the model reasons only over the selected evidence. Final copy follows.

standard_loop.pyStandard tool-calling loop

while not done:
    # full transcript re-sent every call
    response = llm(messages)
    if response.tool_call:
        result = run_tool(response.tool_call)
        messages.append(result)   # raw pages in context
    else:
        done = True

opensymbolic_plan.pyOpenSymbolic

plan = llm.plan(query, schema)      # ~1K tokens, once

evidence = execute(plan, corpus)     # 0 model tokens
# retrieval, filtering, extraction run as code

answer = llm.reason(query, evidence) # ~3K tokens

The introspection boundary

Placeholder. The planner sees summaries and highlights, never raw page content. Context grows by roughly 0.5 to 2K tokens per step instead of the 5 to 30K a raw-data loop re-sends every iteration. An annotated trace goes here.

Annotated trace · placeholder

step 1plan: 1,024 tokensstep 2execute: 0 tokens · 14 documents scannedstep 3evidence: 3 passages · 1,860 tokensstep 4reason: 2,970 tokens

Three blueprints

Placeholder. Three execution patterns cover the workflows that belong on a deterministic engine. Final copy follows.

PlanExecute

Placeholder. Plan once, run as code, reason over the result.

DesignExecute

Placeholder. Design the pipeline, then execute it deterministically.

GoalSeeking

Placeholder. Iterate toward a goal with bounded, checked steps.

Provable answers

Placeholder. When a question has a logical answer, OpenSymbolic hands it to the Z3 theorem prover instead of asking the model to guess. On FOLIO, first-order logic, that reaches 89.2%, near the 96% human bar.

89.2%FOLIO, OpenSymbolic

96%Human bar

Theorem-prover example · placeholder

premises = translate(text)          # model, once
solver = z3.Solver()
solver.add(*premises)
solver.add(Not(conclusion))
result = solver.check()               # unsat → proven

Security by design

Placeholder. The model never touches raw data, so raw data never leaves your boundary. Guarantees listed here in the final copy.

01Placeholder guarantee
02Placeholder guarantee
03Placeholder guarantee

Don't take our word for it. Run a working agent in five minutes.

Tutorials →GitHub →pip install opensymbolicai-core