Skip to main content
Discovery finds you a tool. A workflow tells your agent what to do with it. Hive groups its catalog into 13 workflows. Each one covers a job someone actually asks an agent to do, and carries the sequence of calls that answers it: which tool runs first, what to fall back to when a provider is unavailable, how many billable calls the job should take, and when to stop. When your agent calls search_tools, it gets matching workflows alongside matching tools. Most agents use the workflow.

​
Why this exists

A catalog of 607 tools answers “what can I call”. It does not answer “what should I call, in what order, and when am I done”. Left to work that out alone, an agent tends to make the same mistakes: it calls a tool without the identifier that tool needs, it keeps fetching after it already has the answer, or it hits a rate-limited provider and stops rather than trying the alternative. Workflows carry that knowledge so the agent does not have to rediscover it every session.

​
The 13 workflows

That is 32 routes in total. A workflow has more than one route when the same subject gets asked about in different ways: Security Risk separates a contract check from an approval check from a transaction preflight, because those need different inputs and different tools.

​
What a route carries

Take the spot snapshot route in Market Research, which answers questions like “what is BTC trading at, and is there enough liquid venue coverage?”.
1

Trigger

When the route applies. Here: a current price, market context, or a quick asset check.
2

Required identifiers

What the agent must resolve before calling anything. Here: an exact asset id and a quote currency. Half of all bad tool calls come from guessing an identifier instead of resolving it first.
3

Ordered steps

Each step names a phase, a purpose, a primary tool, and a fallback. The spot snapshot starts by fetching the current observation with get_price.
4

Call budget

maxMaterialCalls caps the billable calls the route should need. Twenty-six of the 32 routes allow 4. The six that allow 2 or 3 are the cheaper jobs: both discovery routes, all three RWA routes, and the chain protocol landscape route in DeFi.
5

Stop condition

What “done” looks like, so the agent returns an answer instead of gathering more.

​
Fallbacks are conditional, not automatic

Every step names a fallback tool, and every fallback carries the condition that permits it:
Use only when the primary is unavailable, gated, rate-limited, stale for the decision, or lacks the required identifier coverage.
Your agent does not silently swap providers to get a nicer-looking answer. When it does fall back, the receipt names the provider that actually served the data, so you can see the substitution rather than infer it.

​
Reading them yourself

search_tools returns workflows at three levels of detail, and defaults to the smallest: Use full only when you need the output schema. On a typical query it returns close to five times the payload of compact, and that context is usually better spent on the problem.
The hive://toolsets resource lists all 13 without a tool call, which is useful when you want to read them once and cache the result.

​
When to skip them

Workflows help when the shape of the job is unclear. They add nothing when it is not. If you already know the tool you want, call invoke_api_endpoint with its name and skip discovery. If every question you ask lives in one category, connect to that category endpoint instead and your agent sees those tools directly, with no discovery step at all.

How Hive works

The discovery loop these routes sit on top of.