How the platform fits together
The layers a conversation passes through, and what each one is responsible for.
Auron is one pipeline with clear responsibilities at each stage. Understanding the shape helps when deciding where a change belongs: whether something is an agent's brief, a signal definition, an automation, or work for your own systems.
Principles
Evidence first. Every output points back at the conversation it came from. A summary can be checked, a signal carries its quote, an action can be traced to what caused it.
Organization-scoped. Everything is created and governed inside one organization. See Organizations and access.
Configuration over code. How an agent behaves, what it notices and what happens afterwards are settings, not deployments.
Integration rather than replacement. Auron holds the intelligence and hands off execution, so adopting it does not mean migrating your operational tooling to it.
The layers
Conversation interfaces
Voice and chat. An agent joins a meeting, takes a call on its own number, or answers in a chat. Transcripts of conversations held elsewhere can be uploaded and treated the same way. See Capturing conversations.
Context assembly
Before and during a conversation, the agent assembles what it needs to be accurate: the record it is about, what previous sessions established, the knowledge stores it has been given, and anything it looks up through its toolkits. This is the layer that decides whether an answer is grounded or generic.
Intelligence extraction
After the conversation, and partly during it: summaries, signals against your catalog, artefacts extracted across the whole transcript, and reports written to a prompt. Governed by your signal definitions and by the AI constitution.
Memory and retrieval
What accumulates. Signals and validated observations build up on the record, with provenance, so continuity survives people changing. Private conversations produce outputs without contributing to shared memory.
Authorization and governance
Enforced throughout rather than at the edge: which organization, which records, which role, which knowledge, which tools, which actions. See Organizations and access.
Actions and events
The output side. Automations run around sessions and can pause for human approval. Webhooks tell your systems what happened so they can act. See Integrating follow-through.
Observability
What was produced, what it cost, and what happened inside a run. See Observability.
Where a change belongs
| You want to change | Change it here |
|---|---|
| How an agent talks or what it asks | The agent's canvas |
| What gets noticed in conversations | Signal definitions, attached to the agent |
| A rule that applies to every agent | The AI constitution |
| What happens after a session | An automation |
| What your other systems are told | Webhooks |
| Who can see what | Roles, teams and record access |