TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 19—21 · FACTUAL · DURABLE

Tool use, reaching the world

The model proposes a tool call and the runtime executes it, so permissions live there.

The idea

On its own, a language model only predicts text. Tool use is how a system lets it reach the world anyway. Each tool — a climate API, a code library, a calculator — is described to the model in words: what it does and what inputs it takes. When the job needs a tool, the model writes a structured request instead of an answer: which tool, with which inputs. The runtime — the software around the model — checks that request, runs the tool, and hands the result back. The model never runs anything itself. It proposes; the runtime executes. Permissions live there.

Why it matters

Tool use is how your existing studio software becomes something the model can consult rather than something it replaces. The calculation is still done by the calculator. The judgment of which tool to reach for, and whether to trust the result, is now shared. You should know where each part of that judgment sits.

See it in the studio

"What is the basic wind speed for Belagavi?" Without tools: a confident number from memory, possibly wrong. With a code-lookup tool described to it: a call, a result with a clause reference, an answer you can check. Same model. The tool gave it somewhere to look.

Watch for this

A model that describes using a tool when no call happened. Check the log. "I consulted the code" can be predicted text like anything else.

Try it

Read the tool description of any tool a model you use can call; many apps show them. Rewrite it badly — vague inputs, no stated purpose — and watch how the model's use of it changes. The description is the brief.

Prove it

Explain how a model that only predicts text ends up "using" a calculator or an API, give the sequence — model proposes, runtime checks and runs, result returns, model continues — and say why the permission belongs at the runtime step.

How it works

Models are trained to write tool requests in a fixed format when given tool descriptions, and to read tool results as ordinary context. The conventions for describing tools are converging across vendors. Because the runtime executes and the model only proposes, the runtime can refuse, sandbox, log or ask a human before anything runs. Put the permission there: a line in the prompt — "never send without asking" — asks; the runtime's gate enforces. Read-only tools run freely. Tools that write, spend or send wait for a human.

What this idea builds on

What this idea opens up

Sources

Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.

Age grows from 11 at the centre to 22 at the edge, and six sectors show the learning strands. Tab into the map and the arrow keys step from idea to idea, following the links where there is one. Enter opens the idea under the cursor, and E reads out its links and the reason recorded on each. Press slash for Search, question mark for the full key list, and Escape to leave. Open Ideas for the complete readable list, including what each idea builds on and what it opens up.

LOGIKA · RBDS AI LAB INDIA
ON-RAMP · AGE 11 · FIRST ENCOUNTERS, NOT GATES — IDEAS · — DEPENDENCIES
DONE
OPENS NEXT
SOLID — STANDS ON