Step 23 — Agents and tools

Goal: let the model use tools and take actions instead of just producing text.

Why this step matters¶

An agent is a model that can call tools, look at the results, and decide what to do next. It thinks, acts, observes, and repeats.

For Sanskrit this is a natural fit, because you built the tools in Step 15. A model that can call a sandhi splitter, a dictionary, and a metre checker is far more capable than one that has to know everything itself.

What to cover¶

1. Function calling. Getting the model to produce valid, structured tool calls reliably. Constrained decoding forces the output to match a schema.

2. The think-act-observe loop. The basic agent pattern.

3. Multi-turn reinforcement learning. Note that plain GRPO from Step 13 assumes a single turn, which breaks here. Multi-turn extensions exist and are an active research area.

4. Memory and context management across many steps.

5. Evaluating agents on task completion, not on whether the response reads nicely. This is a genuinely different kind of evaluation.

A concrete Sanskrit example¶

An agent that helps a student work through a difficult verse:

1. Split the sandhi              → calls your Step 15 splitter
2. Look up each word             → calls a dictionary
3. Identify grammatical forms    → calls your parser
4. Check the metre               → calls your metre checker
5. Propose a translation         → the neural model's own work
6. Verify the translation is
   consistent with the parse     → back to the tools

Steps 1 to 4 and step 6 are deterministic and checkable. Only step 5 needs the neural model. That division is the whole point.