Preparing components...
Skip to content

Interactive explanation of the planning approach. The current product uses an embedded typed graph, not a embedded graph runtime.

Mountain range under an open sky

Sounds right ≠ runs.

Vector search finds tools that sound right. A graph finds the ones your agent can call, in order.

Near-duplicate tools

Three tools score the same. Only one takes the input you hold. The graph knows which.

Missing prerequisites

A PR needs a branch. Nothing in a top-k bag makes one. A path can't have a hole in it.

No order

Similarity returns a bag. Agents need a sequence. The path is the order.

Re-resolving what you know

Agents look up the same repo every run. Switchboard skips steps whose output you already have.

Each tool is safe. The chain is not.

Private data, untrusted input, a way out. Seeing the whole plan first is how you refuse it.

See the guard
fetchfetch
gmailsearch_emails
gmailsend_email

One query. The answer is the plan.

Types and tools alternate, so a path is a call sequence. One typedPathSearch call finds it.

repo.refcreate_branchbranch.namecreate_or_update_filecommit.shacreate_pull_requestpr.number
embedded graph · typedPathSearch
1CALL typedPathSearch({
2 sourceLabel: 'Type',
3 sourceProperty: 'name',
4 sourceValues: ['repo.ref', 'file.path'],
5 targetLabel: 'Type',
6 targetProperty: 'name',
7 targetValues: ['pr.number'],
8 pairwise: false,
9 relTypes: ['FEEDS', 'YIELDS'],
10 relDirection: 'outgoing',
11 maxLen: 8,
12 pathCount: 3,
13 resultLimit: 50
14}) YIELD path, pathWeight, pathCost
15RETURN path, pathWeight, pathCost
16ORDER BY pathWeight ASC

Graph tool retrieval isn't new.

Others walk tool graphs too. Switchboard derives every edge from types, solves the path in one call, and ships as an MCP server.

Search Switchboard

Search the docs, plan a task, or jump to a page