Route intent to the right specialist agent
Stand up a **router agent** that reads the user’s first turn, classifies **intent** in one pass, answers only what fits a tight job description, then **names one specialist** and hands off a tight paste block. Coworkers stop opening the wrong agent and smuggling everything into a mega-prompt.
Plus: three Admin-Agent passes: draft **routing rules** from your agent directory only, stress-test when the router **answers past its lane**, and compare this pattern to **workflows** ([Run a workflow](/tutorials/run-a-workflow)) vs **sequential handoffs** ([Chain steps so agents hand off cleanly](/tutorials/chain-steps-so-agents-hand-off-cleanly)).
| Audience | Admins · Everyone |
|---|---|
| Time | ~12 min |
| Prerequisites | Several live specialists ([Create an agent from scratch](/tutorials/create-an-agent-from-scratch), [Give your agent its job description](/tutorials/give-your-agent-its-job-description)). A discoverability surface ([Build your agent directory](/tutorials/build-your-agent-directory)). Helpful: knowing when splits beat sprawl ([Split an overloaded agent into two](/tutorials/split-an-overloaded-agent-into-two)). Routers organize **after** boundaries exist; they do not replace crisp job descriptions. |
| You'll end up with | One **Routing** agent with explicit **route vs answer** boundaries, a three-column **intent → specialist** cheat sheet stored beside your directory, and house rules for when humans must switch threads themselves. |
When a tutorial shows italic text in quotation marks, it usually mirrors a label or helper string inside Auxot. Product copy changes between releases — if something reads differently in your workspace, trust what you see on screen.
Callouts with a Worth knowing gold accent are meant as must-read context before you move on. Blockquotes that open with Tip are lighter, optional depth.
Why this matters
Once you have more than a handful of agents, every newcomer asks: which one do I open? Failure modes:
- Wrong door: burn tokens re-explaining context to the wrong personality.
- Mega-agent: one overloaded prompt pretends to be legal, engineering, and support (Split an overloaded agent into two).
- Human switchboard: works until Tuesday gets loud.
A router agent is the front desk for intent. It reads the first message, chooses simple FAQ vs specialist territory, and answers with who owns the next step: usually one agent from your directory (Build your agent directory) plus a short context bundle to paste after you switch the picker.
This is not the same shape as sequential handoffs in a workflow (Chain steps so agents hand off cleanly, Run a workflow). Routers cover entry routing. Workflows own step two. Support and triage fleets often embed a router (Build your customer support agents, Build your bug-report triage agents). This lesson is the standalone pattern for any team.
Nothing routes by wish. You write the router job description. You keep the specialist list honest.
Quick start
- List specialists — name and one-line scope per agent from Settings → Agents or your directory (Build your agent directory).
- Mint Routing — Settings → Agents → Add Agent: job description: classify intent; answer only inside
{tiny allowed topics}; otherwise reply with Recommended agent:, Why (one line), and Paste-ready summary: bullets. Never impersonate specialists’ tools (Define a tool policy). Routing tool policy stays narrow. - Dry-run five real asks — paste messages teammates actually send. Routing should name the right specialist from intent alone.
- Publish the cheat sheet — three columns: User intent, First agent, Never send here. Link from your directory landing (Update your agents without breaking the team).
- Teach the switch — humans still pick the agent in the Chat picker. Routing advises. Say that out loud.
Done? Routing live, cheat sheet linked, next busy Tuesday sees fewer wrong-thread opens.
The agent can do that?
1. Directory-grounded routes
Chat → Admin Agent:
Specialist list (paste from directory): […]. Draft Routing job description, max 180 words, include refusal when the question needs tools Routing lacks, markdown.
Why it’s non-obvious: Routers swell until someone forbids tool sprawl. The draft stays disciplined after you paste facts.
2. Over-answer trap
Routing replied with a full legal opinion; specialists exist for legal. Rewrite Routing instructions, two sentences, deferral tone, no new agents.
Why it’s non-obvious: Models hate saying go away. Boundary language fixes morale you enforce.
3. Pattern pick
Volume ~30 chats/day. Compare Routing-only vs workflow entry ([Run a workflow](/tutorials/run-a-workflow)) vs chain ([Chain steps so agents hand off cleanly](/tutorials/chain-steps-so-agents-hand-off-cleanly)). Trade-offs bullets, no vendor fluff.
Why it’s non-obvious: Automation envy erases clarity. Paste volumes you see; keep the reply proportional.
Go deeper
Single-thread honesty
Auxot does not auto-switch the Chat picker. Say so in onboarding. Pins and Slack macros carry the cheat sheet (Connect Slack to your agents).
Tool parity
If Routing lacks GitHub but guesses code answers, attach read-only FAQ context (Add your first context file) instead of widening tools blindly.
Regression hooks
When specialists split again, rerun the dry-run (Catch regressions after you change an agent, same discipline).
Safety
Public ingress gets gamed. Pair with adversarial drills once Routing sits at the edge (Red-team your agents against prompt injection).
Walkthrough
Step 1: Freeze specialist table
Markdown table: Agent | Scope | Tools yes or no. Facts only.
Step 2: Author Routing job description
Sections: Answer, Route, Refuse, Tone. Each at most three bullets.
Step 3: Negative examples
List intents Routing must not solve. Keeps prompts short.
Step 4: Pilot surface
Pick one Slack channel or internal wiki page. Link Routing plus cheat sheet (Build your agent directory).
Step 5: Review monthly
Roster review and stale routes: Audit and clean up your agents. Optional rhythm: Automate weekly checkups on your agents.
Next steps
- → Build an audit narrative from your logs. Prove who routed where when diligence asks for receipts, not only org charts.
- → Tune your router agent when routing misfires. After Routing ships. Harvest misses, classify failure modes, patch job description, retest frozen prompts.
- → Build your agent directory. So Routing copies names humans already trust.
- → Split an overloaded agent into two. When one brain fights itself. Routing organizes after boundaries exist.
- → Chain steps so agents hand off cleanly. For sequential automation. Different problem than pick-the-right-door routing.
- → Run a workflow. When volume deserves structured boards. Routing stays the conversational front end.
- → Build your customer support agents. For a multi-agent example where Routing often appears naturally.
Reference
- Pages in Auxot: Chat, Settings → Agents
- See also: Allocate model spend to teams for internal reporting, Build an audit narrative from your logs, Build your bug-report triage agents, Create an agent from scratch, Define a tool policy