Auxot
Tutorials Docs Pricing Login

Create an agent from scratch

Build your first custom agent — a specialized teammate for one specific kind of work — by chatting through it with the Admin Agent.

Plus: three prompts to paste into Chat — one with your newest agent so it explains its own lane back to you; two back in Admin Agent chat for what to build next and for an agent that reads your other agents' descriptions when you ask it to.

Audience Everyone
Time ~10 min
Prerequisites An Auxot account on any tier. [Say hello to the Admin Agent](/tutorials/say-hello-to-admin-agent) finished — the Admin Agent feels familiar. Helpful: at least one context file ready ([Add your first context file](/tutorials/add-your-first-context-file)), and you've read [Give your agent its job description](/tutorials/give-your-agent-its-job-description) so the idea of an agent's job description isn't new.
You'll end up with One custom agent — named, described, configured, and ready to do one specific kind of work — and a clear sense of how to build the next one.

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

In Say hello to the Admin Agent, we said the Admin Agent’s job is helping you build the agents that handle your business’s actual work. This is that moment. Every business has work that keeps slipping: the research nobody has time for, the follow-ups that fall through cracks, the roles you’ve been meaning to fill. Today you build the first agent that picks one of those up. Not “an AI assistant” but a specific teammate, with a specific job, asked to handle specific kinds of prompts in specific ways.

You describe the work you want offloaded in Admin Agent chat; you prompted first. He asks the sharpening questions and, when you’re ready to commit, lays down the configuration in one motion. By the end of that conversation there’s a new row in Settings → Agents. What’s left is yours: open Chat, pick your new teammate, send a real question, and tighten the Description or files if what comes back misses the lane you had in mind.

Quick start

  1. Sign in: open Auxot in your browser and log in.
  2. Open chat with the Admin Agent: click Chat in the left menu, make sure the agent picker reads “Admin Agent.”
  3. Describe the work you want offloaded: type something like “I want an agent that drafts customer support replies in our brand voice” and press Enter.
  4. Answer the Admin Agent’s questions: he’ll ask what tools the agent should have, any rules it should follow, and what to call it. Two or three exchanges, usually.
  5. Confirm: the Admin Agent proposes the full configuration. Say “yes, create it” (or whatever feels natural). The agent is created on the spot.

Done? Your new agent is now in Settings → Agents, selectable from the agent picker in chat. Send it a message to verify it works.

The agent can do that?

Agent created. Now you prove, in Chat, that someone else would understand its job. These three prompts: one you send to your new agent directly; two you send after switching back to Admin Agent.

1. Have your new agent describe its own job back to you

Flip the picker to the agent you just built (not Admin Agent) and send:

In two or three sentences, describe what you're for and what you'll never do. Then list the three kinds of questions you're best at answering. Be specific.

Why it’s non-obvious: Cheap sanity pass. If what’s written back matches the lane you meant, you’re aligned: if not, Give your agent its job description is nearby. Skip it and folks often discover mismatches halfway through unrelated threads where the teammate has been answering confidently against the wrong shape of job.

2. Ask Admin Agent what to build next

Back on Admin Agent in the picker. New message:

I just built [agent name and one-line job]. Looking at the kinds of work that get done in a [your business — e.g., "small marketing consultancy"], what's the next agent I should build to keep getting more out of Auxot? Recommend one specific agent — name, description, what context files it would need — and tell me why it's the next one.

Why it’s non-obvious: The second agent is where people stall: another whole setup mentally. After you send this prompt, the reply borrows what’s already configured plus how you describe the business and hands back a blueprint tight enough for one more assisted build. You open Admin Agent chat again and walk through creation the same way as this tutorial. You build agents faster when you start from a recommendation than when you start from scratch every time.

3. Build an agent that reviews and tightens your other agents’ descriptions

Still in Admin Agent chat, once you already have several custom agents wired up:

Build me an agent whose job is reading the descriptions and context files of my other agents and telling me where they're vague, where they overlap, and where any description could be tightened. Give it a clear plain name that names the work it does. Give it a clear job description and make it ready to use.

Why it’s non-obvious: You’ve been editing agents by hand whenever something felt muddy. After you send the gray-box prompt and accept the new agent, you have a teammate whose only job is spotting vague overlap. Open that agent’s chat when you want a review (for example after you change instructions). Nothing reviews your other agents on its own. Either you open that agent when you want a review, or you wire a reminder, workflow, or integration (Run a workflow, Connect Slack to your agents territory). A monthly calendar nudge plus ten minutes beats hoping vague overlap surfaces on its own.

Go deeper

Picking a good first agent

A useful agent has a clear sense of what work it takes on, what voice it uses, and what it should never do. Scope can be narrow (one specific recurring task) or broad (a role with several related tasks), both work. The fix when an agent feels generic is rarely “narrow the scope.” It’s “say what the agent should do, in what voice, and what it should never do.” Give your agent its job description has examples of descriptions that work at both scopes.

If a single agent’s description starts feeling like “and also…” over and over, that can be a sign to split it into two more focused agents: not because broad scope is wrong, but because the constraints stop fitting into one description.

How the Admin Agent actually creates the agent

When you confirm and the Admin Agent says it’s creating the agent, what happens behind the scenes: a single internal tool call writes the agent’s name, description, instructions, preferred AI model, attached context files, and tool policy at the same time, atomically. Either everything lands, or nothing does. The agent appears in your list with everything connected.

Troubleshooting

  • The Admin Agent kept asking questions and never created the agent. Sometimes the conversation runs longer depending on how much specificity you brought. If you’ve been at it for five exchanges and want to just create something, say “propose the configuration now and create it.” The Admin Agent will lock down what it has and create the agent.
  • The agent answers in the wrong tone or style. Its description or attached context files need work. See Give your agent its job description.

Variations & edge cases

  • Free tier: unlimited custom agents. The only thing Free tier can’t do is split them across multiple teams (Free is single-team).
  • Visibility scope: on Business and Enterprise, agents can be Team-scoped (only that team sees them) or Org-scoped (everyone in the org). Set this on the agent’s detail page after creation.
  • Editing later is fast. Everything about the agent (description, attached context files, tool policies, name) is editable on its detail page. Auto-save means no Save button.
  • Deleting an agent. Settings → Agents → trash icon on the row. Deleted agents are gone permanently; the chat threads with that agent stay (they’re tied to your user, not to the deleted agent).
  • The Admin Agent itself is locked. You can’t create “another Admin Agent”: it’s the one built-in agent that ships locked. To get an Admin-Agent-shaped agent that you control, build a custom agent with a similar job description and similar context files.

Walkthrough

There are two ways to create an agent in Auxot:

  • The Admin Agent path (recommended). You chat through it. The Admin Agent asks the right questions, writes the agent’s description and instructions for you, and creates it on the spot.
  • The Settings UI path (faster, less guided). Best when you already know exactly what you want. Covered as a single paragraph at the end since it overlaps heavily with Add your first context file (context files), Give your agent its job description (description), and Define a tool policy (tool policies).

Path A: Build it through the Admin Agent

Step 1: Open chat with the Admin Agent

Click Chat in the left menu. Make sure the agent picker at the top reads “Admin Agent.”

Step 2: Describe the work you want offloaded

Rough wording is fine: the Admin Agent’s job is to ask the questions that turn what you said into a real agent. Something like:

  • “I want an agent that drafts customer support replies in our brand voice.”
  • “I need an agent that reads our weekly leadership minutes and tells me which commitments have gone quiet.”
  • “I want an agent that reviews contracts for compliance with our standard terms.”

The more concrete, the better. ”I want a marketing agent” is too vague: what kind of marketing? Who for? What does it produce? The Admin Agent will ask if you start vague, but the conversation goes faster when you bring the specifics with you.

Step 3: Answer the Admin Agent’s questions

This usually takes two or three exchanges. The Admin Agent works through:

  • What it should do: the agent’s actual job. The Admin Agent will sharpen what you said in Step 2 into a one-line description.
  • What tools it should have: does it need to read external websites? Send Slack messages? Look up data in your CRM? The Admin Agent shows what’s available and helps you pick.
  • What it should never do: hard rules, escalation paths, things that would get a new hire fired on day one.
  • What to call it: a clear name that anyone scanning the agents list will recognize.

Tip: Treat each exchange like you’re hiring a new employee. What would the new hire need to know on day one? What would get them fired? That’s the level of specificity that produces agents that actually work. If the Admin Agent proposes something that doesn’t fit, say so: it’s a conversation, not a form.

Step 4: Confirm the configuration

When the Admin Agent has enough, he proposes the full configuration: name, description, instructions, tools, anything else relevant. Read it. If it looks right, say ”yes, create it” (or anything similar: he picks up on intent, not magic words). The agent is created on the spot. He’ll tell you it’s done and where to find it.

Step 5: Find your new agent and test it

Two places it now lives:

  • Settings → Agents: your new agent appears in the agents list. Click it to see and edit anything.
  • Chat picker: the new agent is selectable in the agent picker at the top of the chat window.

Open chat with your new agent and try it on a real task. If the answers feel off, head to Give your agent its job description to refine the description and attached context files.

Tip: If you want this agent to always use a specific model (Claude, GPT-4, your local GPU), tell the Admin Agent when you’re setting it up: it sets the preferred model at creation time.

Path B: Build it through Settings → Agents

Click Settings → Agents → +Add Agent, fill in Name, Description, and Team, and click Create. The agent appears on its detail page where you wire up the rest: description (Give your agent its job description), context files (Add your first context file), and tool policies (Define a tool policy). The agent is chattable from the moment you click Create: it just won’t have specialized knowledge or instructions until you wire those up. Faster than Path A, but only worth it when you already know exactly what you want.

What’s next

  • → Add your first context file. If your new agent doesn’t have any business knowledge yet, this is the next step.
  • → Give your agent its job description. Refine the description as you learn what your new agent does well and where it misses the mark.
  • → Define a tool policy. Give your agent real-world capabilities (search, code, internal APIs) on top of its context files.
  • → Connect a cloud AI model. When the agent exists but the picker still lacks a model you trust, wire OpenAI or Anthropic so answers have a real engine behind them.

Reference

Browse all tutorials