Collect real APIs
Start with concrete schemas, parameters, constraints, and outcomes. The diversity of the catalog is the training surface.
From catalog to call tree
Search for the call, not the line.
Thousands of tools become useful when an agent can compare solution paths instead of scanning a catalog one item at a time. Build the tree, inspect the alternatives, then export the winner.
Five moves. One searchable system.
The original sequence is preserved here as a connected pipeline. Each move changes what the next stage can learn, compare, or retrieve.
Generate instructions
Turn each tool contract into task-facing guidance: when to call it, what evidence it needs, and what it returns.
Annotate solution trees
Keep alternatives visible. A branch records the call, its place in the path, and why it competes.
Finetune on paths
Train on successful routes and the decisions between them, not isolated tool names detached from tasks.
Search for the call
At inference time, rank plausible next calls and prune weak branches instead of walking the catalog linearly.
Give the task a search tree.
Define the search space
Format: name | comma-separated task words | reliability 0-1 | cost 1-5
Ranked call paths
4 candidate branches
Winning path
Weather.now → SMS.send
Matches the task's earliest high-signal intents with strong reliability and low call cost.
“Collect APIs. Generate instructions. Annotate solution paths with a tree.”
The instrument above does not pretend to call those APIs. It makes the post's search object visible: a task-conditioned set of competing paths that can be ranked, pruned, saved, and exported.