Open-source AI teammates in Slack, Teams, Discord, Telegram and Mattermost. Self-host it, with any agent harness and any model.
Decision routing (auto-route): setting it up and what rooms do
Auto-route lets Switch pick the agent that answers a message that tags nobody.
In a room with !auto-route on, a person's untagged message that no follow-up
or assigned agent took is sent, with the room's recent context, to a decision
service, which names one of the room's agents or nobody. The service itself
(decision/, image switch-decision) is documented in decision/README.md;
this page covers how a server and its workspaces choose one, and what a room
does when there is none or it fails.
Turning it on for a server
Auto-route needs two feature flags, follow_the_conversation and
decision_routing, in FEATURE_FLAGS_ENABLED. With the flag off nothing is
ever sent to a decision service, !auto-route says the feature is off, and
the room toggle shows it unavailable.
The server default
The server's own decision service is the default for every workspace that has not chosen one:
| Setting | Meaning |
|---|---|
DECISION_ENDPOINT_URL |
The decision service. Its base URL (http://host:8230) or its /decide URL both work. The bundled Helm chart (decision.enabled) and the compose decision profile set it. |
DECISION_API_KEY |
Sent as the bearer token, when the service requires one. |
DECISION_TIMEOUT_SECONDS |
How long one decision may take (default 5), for every workspace's backend. |
DECISION_BEDROCK_ENABLED |
Whether a workspace may choose Haiku on Amazon Bedrock (default off). Bedrock runs in the server's decision service on the server's own AWS credentials, which is why the operator has to allow it. |
DECISION_PENDING_MAX_AGE_SECONDS |
How long a message may wait for a decision, across a restart or an outage, before it is dropped rather than answered late (default 120). |
DECISION_SWEEP_CONCURRENCY |
How many workspaces' waiting decisions each process's sweep settles at once (default 4, 1 to 64). |
With no DECISION_ENDPOINT_URL, there is no server default: auto-route is
unavailable in every workspace until its admins choose a backend, and Haiku is
not offered (it runs in the server's decision service).
The model behind the server default
The bundled Helm chart runs Strands Decider 4B (AWS's Strands Decider, the checkpoint built on Google's Gemma 4 E4B) on one NVIDIA GPU, beside the decision service. It was chosen on a benchmark of 100 routing scenarios in four rooms with overlapping agents, a third of them messages nobody should answer, run on one NVIDIA L4:
| Model | Correct | Wrong picks | Stayed quiet when it shouldn't have | Median |
|---|---|---|---|---|
| Strands Decider 4B | 92 | 2 | 6 | 0.27 s |
| Qwen3.5-4B, 4-bit (llama.cpp) | 88 | 9 | 3 | 0.16 s |
| Claude Haiku 5.5 (40-scenario subset, for scale) | 38/40 | 0 | 2 | 1.5 s |
The four-point gap is within noise at 100 scenarios; the wrong picks decide it. A wrong pick is the failure people notice, an agent jumping into a conversation that was not for it, and most of Qwen's were exactly that (answering a message meant for a named person, or people talking among themselves). Strands Decider's misses are the quiet kind: it stays out, its confidence usually just under the 0.6 threshold.
- Switching to Qwen is one chart value,
decision.modelServer.kind: llamacpp, which runs Qwen3.5-4B on llama.cpp's CUDA server instead (smaller: 8 GiB of RAM rather than 24). Local compose uses Qwen by default, because it runs on a CPU; seedocs/old/LOCAL_DEVELOPMENT.md. - Licence. Strands Decider is Apache-2.0, but its 4B checkpoint is built on
Google's Gemma 4 (
google/gemma-4-E4B-it), whose terms apply to whoever deploys it. The weights download from Hugging Face at deploy time, into the deployment's own volume; Switch does not redistribute them. - Hardware. One NVIDIA GPU with 24 GB (an L4) and a node with 32 GiB of RAM (an AWS g6.2xlarge): the 16 GB of bf16 weights pass through memory while they load. With 16 GiB the model took about 10 minutes to start.
Each workspace's choice
A workspace's admins (its owners and admins, and the server operator) choose
in Switch Console → Admin → Decision routing. Switch stores one row per
workspace (decision_settings); no row means the server default.
- Server default: whatever the server is configured with. "Use server default" deletes the workspace's row.
- None: auto-route is off in the workspace's rooms, even with a server default. Rooms behave as they do without it.
- Self-hosted endpoint: the workspace's own decision service, a URL and an optional key sent as its bearer token. That workspace's room context goes there, by its admin's choice; no other workspace is affected.
- Haiku (Anthropic API): the workspace's Anthropic key, sent per request to the server's decision service, which calls Anthropic with it.
- Haiku on Bedrock: offered only where the operator set
DECISION_BEDROCK_ENABLED; no key is stored. - Confidence threshold: between 0 and 1, sent to the service as the confidence it needs to name an agent. Empty leaves it to the service.
A saved key is encrypted under the server's keys (SECRET_KEYS, re-encrypted
on rotation like every other stored secret) and only its last four characters
are ever shown. A key the server can no longer decrypt still loads, shown as
set but unreadable, so an admin can save a replacement; until then that
workspace's decisions fail (see below).
A save applies to the workspace's next decision with no restart: the process that saved it at once, the others within a few seconds.
Test runs one fixed sample decision (three made-up agents and a deploy question, nothing from any room) against the values entered, saved or not, and shows the pick, the confidence and the time it took, or why it failed: not reachable, timed out, key refused, an answer that is not a decision, a backend the service cannot run, or nothing to test.
The same settings are served by the gateway API, for that workspace's admins
only: GET, PUT and DELETE /gateway/decision-settings, and
POST /gateway/decision-settings/test.
In a room
!auto-route on and the auto-route toggle on the room page (Console and the
gateway, PUT /gateway/rooms/{id}/auto-route) set the same switch. Turning it
on is refused, with the reason and who can fix it, where it cannot act: the
flag is off, or the workspace resolves to no backend. A room's detail reports
auto_route: {on, available, reason, fallback_agent, fallback_agent_name}.
Who can change it
- Turning auto-route on or off takes write access to the room, the same as
!follow-upsand!assign-agent: its owner, a workspace admin, or anyone on a room everyone may write to. In chat, anyone who can post in the room may send!auto-route on|off. In Console and on the gateway's room page the toggle follows the room'scan_write, and is shown disabled, with the reason, to someone who can see the room but not write to it. The API answers 403 without write access. - Naming or clearing the triage agent takes the room's owner or a
workspace admin (the room's
can_manage), because it chooses an agent that gets messages nobody addressed to it.!auto-route fallback @agent|offfrom anyone else is refused with a reply saying who can; the picker in Console and on the gateway is read-only for them ("Only the room's owner or a workspace admin can change this"); and aPUTcarryingfallback_agentanswers 403. Anyone may see which agent it is. - Choosing the workspace's decision backend stays with workspace admins (above).
When there is no backend
Nothing changes from a room without auto-route: tags, aliases, roles, direct rooms, follow-ups (in threads too) and the assigned agent all work, and no message content is sent anywhere.
When the backend fails
A timeout, a refused connection, an answer other than 200, or an answer that is not a decision (including one naming an agent that was not a candidate) is a failure. A decision of nobody is not.
- The message goes to nobody, and the server log names the failure.
- Switch posts "Auto-route is unavailable right now; tag an agent to reach
it." in the room, at most once an hour per room. The time is kept in the
room (
decision_notice_at), so restarts and several processes don't repeat it. - Triage agent: a room can name one of its agents with
!auto-route fallback @agent(or from the toggle). While the backend fails, the message goes to that agent instead, through the same addressing-policy and budget checks as any routed message, markedrouted_reason: "fallback"so it knows why it got it. It cannot be the room's assigned agent, which already gets every untagged message before auto-route is asked.!auto-route fallback offclears it. - Thread replies too. In a room with auto-route on, a reply in a thread is decided like any other untagged message: the follow-up rule that sends it to the agent that spoke there last is skipped. So while the backend is failing, a reply in an agent's thread that doesn't tag it goes to the room's triage agent if one is named, and otherwise to nobody, with the notice above; in a room without auto-route it would have reached that agent. People reach it meanwhile by tagging it, which works as always.
Across a restart
A message waiting for a decision is recorded with the message
(pending_decisions) and decided by whichever process gets to it first: the
one that wrote it at once, any other on its sweep every few seconds, or the
same one after a restart. Each is delivered at most once. One still waiting
after DECISION_PENDING_MAX_AGE_SECONDS is dropped and logged. The sweep
visits only the workspaces with a message waiting (one query across
workspaces), up to DECISION_SWEEP_CONCURRENCY of them at a time, so a
workspace whose backend is slow or failing holds up only its own messages.
What is reported
Every decision is logged with its room, candidate count, choice, confidence,
threshold, latency and backend, never the message. Product telemetry, when it
is on, counts decisions and failures without any room, agent or message
identifier: decision_made and decision_failed in
docs/old/telemetry-events.md.