Hyreflow Agent
Judgment at list scale, on the rows nobody wants to read.
Sometimes the work is not a lookup, it is a decision repeated five hundred times. Is this company really in scope? What does this job ad actually want? Our own hosted model handles that row by row with nobody in the loop. You pay for what it reads and writes, not per call.
It runs in the middle of a pipeline, wherever a step needs judgment rather than a database.
things you can ask for
- “Score these 400 companies from TheirStack against my ICP”
- “Pull the real requirements out of these job ads and turn them into a brief”
- “Write a first line for each of these before they go into SmartLead”
- “Research these firms and tell me which are worth calling”
works well with Hyreflow Agent
- TheirStack
the long company lists worth scoring rather than reading
- Hyreflow Native
raw job ads in, structured requirements out
- SmartLead
personalisation at the scale volume sending needs
- Firecrawl
the pages the agent reasons over
Ask for the outcome and your agent composes the run itself. Or start from a skill, a whole pipeline it already knows end to end.
browse all skills →What your agent reads before it touches Hyreflow Agent
Every tool ships with a written playbook, and the agent loads it before the first call. Auth, rate limits, what each call costs, which actions need your approval, and the mistakes worth avoiding. It is the difference between an agent that knows the tool and one that guesses at it.
Show the raw playbookwritten for the agent
HOW in lib/hyreflow_agent.py.
What it is
A hosted model you invoke to do model work on a row or a task, with two entry points:
infer(prompt, schema=…)— a plain model call, no tools. Classify, extract, score, write. Cheap.research(prompt, schema=…)— agentic: the model plus a toolbelt of Hyreflow adapters in a tool-calling loop, for open-ended research, signal extraction and synthesis.
Pass a JSON schema to force structured output. {{placeholders}} in a prompt are filled from row data by
the caller, which is what makes this usable across a whole list.
Toolbelt (what research() may call)
exa_search— neural/semantic web search: use it first, it has the best recall on an open-ended question.serper_search— Google SERP: forsite:-scoped or exact-string lookups, and to fill gaps Exa leaves.firecrawl_scrape— scrape ONE known URL to markdown, only once you know the page.
Kind: Hyreflow Native
First-party, no BYOK — you never hold a model key. Billing is token-metered: you're charged on the
tokens the call actually consumes, so there's no flat per-call price. infer is materially cheaper than
research, which pays for both reasoning and the tool calls it makes.
When to reach for it
Bulk qualification, personalization at list scale, and company/person research that needs judgment rather than a lookup. For one-off judgment inside a conversation, let Claude reason directly — this Native exists for the headless, per-row case.