How Hive works - Hive Documentation

This page explains how Hive is put together, for technical evaluators and developers. It follows data from your tools, through detection and planning, to a decision and an action — and shows where each safety control sits.

The big picture

Component	What it does
Web app	A client-side React app. It holds no business logic of its own and talks to the API over HTTPS.
API	Verifies every request’s sign-in session, resolves the workspace, and serves reads and writes.
Workers	Run background jobs: collecting data, detecting signals, planning, executing approved actions, running workflows and agents.
Postgres	The system of record. Row-level security is forced on workspace tables; vector search powers knowledge retrieval.
Job queue	Delivers signed jobs at least once, with per-workspace idempotency keys and fairness caps so one busy workspace cannot starve another.
Redis	Caching, rate limiting, scheduler locks and webhook de-duplication.
Connector platform	Handles OAuth sign-in with your tools, stores their tokens, and makes API calls on Hive’s behalf.
Model providers	Language, embedding and image models. See AI and models.

The loop: sense, plan, decide, act, learn

Hive’s core is a loop that runs continuously for every workspace.

Collect

Every hour, and immediately after you connect a tool, workers read from your connected accounts through the connector platform. Some providers also push changes in real time through signed webhooks. Files from Google Drive, Dropbox, OneDrive and SharePoint are synced into the knowledge base with their permissions.

Sense

Detectors compare what they read with learned baselines. A statistical detection only fires when it is well outside the norm on at least two of the last three readings, so a single spike never raises a signal. Built-in detectors also watch email threads and specific business processes. See Signals.

Plan

For a signal worth acting on, the AI drafts a plan: a sequence of typed actions with structured parameters, grounded in evidence from your tools and the Business Brain.

Decide

A deterministic policy engine — code, not a model — checks the kill switch, the action’s autonomy rung, its blast-radius tier, confidence, spend caps, rate limits and error budgets. The outcome is run, ask a person, or simulate. See Safety and control.

Act

An isolated executor, the only component that can use your credentials, carries out the approved action and records a receipt. Reversible actions record how to undo them.

Learn

Approvals, denials and outcomes feed execution memory. New behaviours start in shadow and only take effect when a person activates them. See Learned behaviours.

Requests from people

Ask Hive, the workflow builder and Agent Studio follow the same pattern. The API verifies your session and role, gathers the context the request needs, and calls a model. Anything that would change something in your tools becomes a typed proposal and goes through the same policy engine as the loop.

Workflows and agents at run time

A workflow is a graph of steps compiled and validated at deploy time into an immutable release. Each run executes that release step by step on the workers, persisting every step so a run survives restarts. See Workflow builder.
An agent is a saved, versioned configuration — instructions, knowledge, tools and governance limits — that runs with a budget of time, model calls and actions. See Agent Studio.
Both can be started by people, schedules, forms or other agents, and every external action they propose passes the policy engine.

Where the safety controls sit

Control	Where it is enforced
Workspace isolation	Every query, and forced row-level security in Postgres
Authentication	API, on every request
Roles and approvals	API and policy engine
Kill switch, caps and breakers	Policy engine, before every action
Credential access	Executor only; never the model
Untrusted content	Marked as data in every prompt
Tool integrity	MCP tool fingerprints pinned at connect time
Audit	Written for every executed, approved or undone action

Security overview

The protections in plain language.

Infrastructure

Providers and regions.

Developer overview

What you can build on Hive.

Key concepts

The vocabulary of Hive.

​The big picture

​The loop: sense, plan, decide, act, learn

​Requests from people

​Workflows and agents at run time

​Where the safety controls sit

​Related