Spring Voyage is domain-agnostic: the platform supplies the primitives, and packages supply the expertise. People and agents are peers on one collaboration surface — an agent can ask a person for judgment as readily as a person can ask an agent for work.
Eight ideas carry the platform. The concept docs go deeper on each.
An addressable AI entity with a mailbox, an execution config, and a role. It is the atomic unit of work: every message that needs reasoning is delivered to an agent.
Agents are reachable by address, and they reach each other — and you — the same way.
A first-class member of a team, not an observer. A person joins as an Owner, Operator, or Viewer, and can ask the team questions, answer the questions it asks back, approve work, and intervene part-way through a task.
The identity you wear in one collaboration is a Hat. The same person can wear a different Hat on every team, each with its own display name and role.
A team. Agents and people are organized into units that work on a domain together, each member with its own role, responsibilities, and expertise.
A unit exposes a deliberate face to the outside while preserving deeper access for the people permitted to have it. Units compose, so a team can itself be a member of a larger organization.
A collaboration is the active shared space a set of participants works in — where tasks, decisions, plans, and execution happen.
An engagement is the enduring version of it: the ongoing shared context you return to, rather than a chat you restart.
A pluggable bridge to an external system — GitHub, Slack, email. Agents act in the places where the work already lives, instead of in a separate world you have to keep in sync.
People are reachable the same way: a message addressed to a person is routed to whichever channel they actually read.
The platform is deliberately ignorant of your domain. Expertise arrives as a package — a bundle of agent and unit definitions, skills, and templates.
Install a package and you get a working team, not a blank canvas. spring package install <name>
How far an agent may go in acting without being asked — from Passive, through Attentive and Proactive, to Autonomous.
Unit policies hold the line: allowed and blocked actions, cost limits, and approval requirements.
Every agent and connector emits a structured activity stream — reasoning, decisions, tool calls, errors — observable in real time by humans and by other agents.
Each event carries a severity and a correlation ID that links it to related events across the system. Where an LLM was involved, it carries its cost.
