Core Concepts
The platform vocabulary, defined precisely. Each term has a one-line definition and a link to its deep-dive page.
#The model
Every concept on the platform fits one mental model. An agent perceives through connectors, learns procedures through skills, carries context through memory, and takes outbound action through actions. That pipeline runs as a sequence — Ingest, Extend, Equip, Remember, Act — and the platform wraps it in five layers so agents stay accountable end to end.
- Deploy — build and run production agents and multi-agent teams.
- Protect — runtime enforcement: approval gates, kill switches, input and output validation.
- Observe — step-level traces, anomaly detection, metrics, and cost tracking.
- Connect — plug agents into any data source, tool, or endpoint.
- Govern — just-in-time authorization, the policy engine, role-based access control, and the audit trail.
The terms below are the building blocks of that model. For the full picture, see the Introduction.
#Agents
An agent is an autonomous unit that reasons over a model, calls tools, and produces output. You define an agent as a typed module: instructions, a model, and the connectors, skills, and actions it can use. Agents can run alone or coordinate as a team, where one agent delegates sub-tasks to others.
#Runs
A run is one isolated execution of an agent. A run is queued when a trigger fires, transitions to running while the agent reasons and calls tools, and ends in a completed or failed state. Each run is independent — its execution is sandboxed, so concurrent runs never share state, and a failure in one run does not affect another.
#Connectors & MCP
A connector is how an agent perceives and acts on a system. A connector exposes typed tools — read a database, call an API, write a file — and the agent invokes them during a run. Connectors speak the Model Context Protocol (MCP), an open standard for governed tool access, so any MCP-compatible source plugs in without bespoke glue.
#Skills
A skill is a versioned, composable procedure an agent can load on demand. A skill is a directory with a definition file describing what it does and when to use it, plus an optional body of methodology, scripts, or reference material. Skills keep procedural knowledge out of the prompt and make it reusable across agents.
#Actions
An action is a structured, typed output an agent emits to drive downstream work. An action has a schema, so its payload is validated and fully typed when your application or a webhook receives it. Actions are how an agent’s reasoning turns into real effects in your systems.
#Memory
Memory persists context across runs. Where a run is ephemeral, memory lets an agent carry forward what it learned — prior findings, decisions, state — under policy-based access control. Memory is managed by the platform rather than wired by hand.
#Triggers
A trigger is what starts a run. A trigger can be a schedule, an inbound event, or a webhook. Triggers connect external conditions to agent execution, so a run begins when something happens rather than only when a person asks.
#Policies & governance
Every tool call and every model output in a run passes through the governance gateway. A policy decides the outcome: allow, deny, or escalate to a human. The gateway is deny-by-default — nothing is permitted unless a policy explicitly allows it. Policies are scoped by role through role-based access control, so the same agent has different permissions depending on the context it runs in.
Read the Governance & Policies guide →
#Human-in-the-loop & JIT
When a policy escalates, the run pauses and a human is brought in to approve or reject the pending action — this is human-in-the-loop. The grant a person issues is just-in-time authorization: scoped to that action, in that run, at that moment, rather than a standing permission. A kill switch can terminate a run immediately if something goes wrong.
Read the Human-in-the-Loop guide →
#Observability & audit
Observability is the live view of execution: every run is streamed step by step into the control plane, with traces, metrics, and cost. The audit trail is the durable counterpart — every governance decision is recorded immutably with the context needed to reconstruct what happened, which is what makes the platform suitable for regulated environments (SOC 2 Type II, ISO 27001, HIPAA, GDPR).
Read the Observability guide · Read the Audit & Compliance guide
#Where to go next
- Quickstart — Scaffold a typed multi-agent project and run your first agent locally in minutes.
- Agents — Define single agents and multi-agent teams with connectors, skills, and actions.
- Governance & Policies — Deny-by-default policy evaluation, roles, escalation, and the audit trail.
- Connectors — Give agents governed tool access to data sources and systems over MCP.