Routing & Sessions

When a user sends a message, the orchestrator needs to decide which agent handles it. Routing rules in agent.yaml control this behaviour. Sessions track the ongoing conversation state between a user and an agent.

Routing modes

Selu supports two routing modes:

Inline mode (default)

In inline mode, the orchestrator evaluates every incoming message against all installed agents and picks the best match. This is the default when a user has multiple agents installed.

routing:
  mode: inline
  triggers:
    - intent: weather_query
    - mention: "@weather"
  priority: 10

The orchestrator scores each agent based on trigger matches and priority. The highest-scoring agent handles the message.

Dedicated mode

In dedicated mode, the agent owns an entire channel or conversation thread. All messages in that context go directly to it — no routing evaluation is needed.

routing:
  mode: dedicated

This is useful for single-purpose bots or when the user explicitly starts a session with a specific agent.

Triggers

Triggers help the orchestrator match messages to agents in inline mode.

Trigger	Description	Example
mention	Matches when the message contains a specific @-mention.	@weather
intent	Matches when the orchestrator’s lightweight classifier detects a semantic intent.	weather_query
keyword	Matches when specific keywords appear in the message.	forecast, temperature
regex	Matches a regular expression against the message text.	\bweather\b

Triggers are evaluated in order. The first match wins for that agent, and its priority score is used in ranking.

Note

Intents are classified by a small, fast model before the main LLM call. This adds negligible latency but gives you semantic matching beyond simple keyword lookups.

Priority

When multiple agents match, priority breaks the tie. Higher values win. The default is 0.

routing:
  priority: 10   # higher = more likely to be selected

Sessions

A session is a stateful conversation between a user and an agent. Sessions track:

• Message history — The rolling window of messages sent to the LLM.
• Memory context — Long-term facts stored across sessions (when enabled).
• Active capability state — If a capability uses multi-step flows, the session holds intermediate state.

Session configuration

memory:
  session_ttl: 3600       # seconds before an idle session expires
  max_history: 50         # max messages retained in the context window
  long_term: true         # persist facts to vector-backed long-term memory

Short-lived sessions

For stateless utility agents (e.g. unit converter), use a low session_ttl and disable long_term:

memory:
  session_ttl: 300
  max_history: 10
  long_term: false

Persistent sessions

For personal assistants that should remember the user over time:

memory:
  session_ttl: 86400
  max_history: 100
  long_term: true

Session handoff

When an agent delegates to another agent using delegate_to_agent, the orchestrator can either create a new session for the target agent or continue the existing one. This is controlled by the delegate_to_agent tool’s new_session parameter.

Next steps

Now that you understand routing, head to Build Your First Agent to put it all together.