docs/loop/reason.md

Reason — the third stage of the convergence loop

Reason is where Unisona decides what to do. With a fresh observation and the relevant memory in hand, this is the stage that plans: what the user actually wants, which tools would help, and how to put together a useful answer.

What it means in Unisona

Reasoning is done by a large language model — and in Unisona the model is interchangeable. Claude, GPT, Gemini, Grok, or a local on-device model can each plug into the same loop; the system routes to whichever is the best fit (and can fall back if one is unavailable). The loop never assumes a specific brand of intelligence.

The model doesn't just produce text — it decides which capabilities to invoke. Should it search the web? Read a file? Generate a document? Look at a repository or market data? Those decisions are made here, and carried out in the next stage. Reason also sets the standard for the answer: deliver substance first, make reasonable assumptions rather than interrogating you with forms, and mark real gaps honestly instead of bluffing.

Where it sits in the loop

Reason turns the context assembled by Remember into a plan. When that plan calls for doing something in the world, it hands off to Act.

The convergence loop

Observe → Remember → Reason → Act → Verify → Converge — then back to Observe.

  • Observe — take in the world
  • Remember — recall what matters
  • Reason — decide what to do (you are here)
  • Act — do it with real tools
  • Verify — check it against reality
  • Converge — record the outcome and improve

See the North Star briefing for the whole picture.