Tutorial 24

Audit and clean up your agent fleet

Why this matters

Agents accumulate. You build one for customer support, one for research, one for contract review — within three months you have eight, two of them haven’t been used in weeks, two of them are doing each other’s jobs without you realizing it, and one of them is quietly answering questions it was never meant to handle.

Most people don’t audit their agents because the audit work is tedious — read each Description, recall what you actually use it for, compare against your other agents, decide what to do. The trick is to delegate the audit. The Admin Agent already knows everything about your account; you build a dedicated Agent Reviewer that does the work for you, and run it monthly. Today, build that agent. By the end, your fleet is cleaner and your maintenance is on a clock.

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. Ask him to build the Agent Reviewer — paste this: “Build me an agent called ‘Agent Reviewer.’ Its job is to read the Descriptions and context files of all my other agents, then tell me where they overlap, where any description is vague, where I’m missing an obvious agent, and which agents haven’t been used in 30 days. Make it ready to use.”
4. Switch to chat with Agent Reviewer — pick it from the agent picker.
5. Ask for the audit — “Run the audit. Be specific about which agents to consolidate, delete, or sharpen, and tell me what’s missing.”

Done? You have a punch list of recommendations from Agent Reviewer. Pick the obvious wins and act on them. The rest can wait for next month’s review.

The agent can do that?

Audit work scales with how often you do it — the longer between audits, the messier the surface area gets. These three prompts make audit work into something the Admin Agent does for you.

1. Have the Admin Agent compare two agents and recommend keep, merge, or split

Open chat with the Admin Agent and try this:

Look at my "[agent A]" and "[agent B]". Where do their jobs overlap? Where do they leave gaps? If you were running my business, would you keep both, merge them into one, split one of them into two, or delete one? Be specific about which fields to change in each case.

Why it’s non-obvious: Most people audit agents one at a time. The pairwise comparison surfaces overlap and gaps that single-agent reviews miss. The Admin Agent doesn’t just say “you have two marketing agents” — he tells you whether to merge them and which Description to keep, or to split one of them along a specific seam.

2. Have the Agent Reviewer post a monthly report to Slack on a schedule

Once the Reviewer is built, you don’t want to remember to run it. Schedule it. Open chat with the Admin Agent:

I want my Agent Reviewer to run on the first Monday of every month and post the audit to my Slack #operations channel. What do I need to set up — a workflow, a scheduled job? Walk me through it.

Why it’s non-obvious: The Admin Agent connects three things you’ve already configured (Tutorial 13’s workflows, Tutorial 16’s Slack integration, the Reviewer agent you just built) into one recurring habit. You stop being the dispatcher; you become the recipient. Scheduled audits catch drift that monthly-but-manual audits miss because they actually happen.

3. Have the Admin Agent recommend the agents you don’t have but should

Inverted-usage move. Once your existing agents are clean, ask:

Looking at my agent fleet and what kind of work my business actually does — [briefly describe] — what agents am I missing? Which of these missing agents would be the highest-leverage to build first, and what would its description and context files look like?

Why it’s non-obvious: Audits usually focus on what to remove. This inverts — what should you add? The Admin Agent reads your existing fleet as evidence of what you’ve prioritized, then names what’s conspicuously missing. Pairs well with Power Move 1: clean up first, build the next one second.

Go deeper

How often to audit

Monthly is enough for most fleets. Weekly is overkill unless you’re building agents at a fast clip. Quarterly lets too much drift accumulate. The Reviewer’s job is to make monthly cheap — once it’s built and scheduled, it costs you ten minutes to read the report and act on the obvious wins.

What “drift” looks like in practice

Three patterns to watch for:

• Overlap drift — two agents whose Descriptions are close enough that you can’t quickly tell which one to pick. Solution: merge or split along a clear line.
• Scope drift — an agent that started narrow (“draft customer replies”) and accumulated extras (“and also analyze sentiment, and also draft pitch emails, and also…”). Solution: trim the Description back to original scope, or split.
• Disuse drift — an agent you built once and haven’t used in 30+ days. Solution: delete or rename with an [archived] prefix so it falls out of your default picker.

When to consolidate vs delete vs sharpen

• Delete when the agent has near-zero use AND its work is fully covered by another agent.
• Consolidate when two agents are doing similar work but you’d lose nuance by collapsing them. Pick one’s Description as the base, fold the unique constraints from the other in, then delete the absorbed one.
• Sharpen, don’t delete when the agent is used but answers feel generic. The fix isn’t deletion — it’s a tighter Description and better context files (Tutorial 05).

The Agent Reviewer is itself an agent

The same governance applies: it has a Description, it can have context files, it can be sharpened over time. After your first audit, the obvious next step is to ask the Admin Agent to refine the Reviewer’s Description based on what was useful in the report and what you ignored. Iterate on the auditor the way you iterate on any agent.

Troubleshooting

• Agent Reviewer’s report is generic and not useful. Its Description probably doesn’t include enough about your business or your fleet’s structure. Edit its Description to mention your team size, your industry, and what kinds of work your agents do. The report will sharpen on the next run.
• Agent Reviewer keeps recommending the same things month after month. That’s not a bug — it’s signaling that you’re not acting on its recommendations. Either act on them, or update its Description to acknowledge constraints you’re working around (“don’t recommend deleting these two agents, they’re for a project that’s still active”).
• You don’t know which agents are unused without checking each one. The Admin Agent can list this — ask: “Which of my agents haven’t had a conversation in the last 30 days?” He uses the audit logs (Tutorial 12) to surface it.

Variations & edge cases

• Free tier: Agent Reviewer works fine on Free tier. The 30-day usage check uses your audit logs which are available on every tier.
• Multi-team setups (Business+): if your org has multiple teams (Tutorial 20), each team has its own fleet. Either build a per-team Agent Reviewer scoped to that team, or one org-scoped Reviewer that audits everything.
• The Admin Agent itself. Don’t audit the Admin Agent — he’s locked, you can’t change his Description, and his “scope” is system-wide by design. Skip him.
• First audit always finds the most. Subsequent audits find less because you’ve already cleaned up. Don’t be surprised when month two surfaces three issues instead of fifteen — that’s the system working.

Walkthrough

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: Build the Agent Reviewer

Paste this:

Build me an agent called "Agent Reviewer." Its job is to read the Descriptions and context files of all my other agents, then tell me where they overlap, where any description is vague, where I'm missing an obvious agent, and which agents haven't been used in 30 days. Give it a clear job description, attach any relevant context files, and make it ready to use.

The Admin Agent proposes the configuration (name, description, context files, tool policy). Confirm by saying “yes, create it” or similar. The agent appears in Settings → Agents.

Tip: If you have a context file that describes your business or your team’s work, attach it to the Reviewer. Its audit will be sharper because it knows what your agents are supposed to be doing for you, not just what their Descriptions say.

Step 3: Switch to the Agent Reviewer

In the chat, click the agent picker and select Agent Reviewer. The chat thread is fresh — no history with this agent yet.

Step 4: Run the audit

Paste this:

Run the full audit. List every custom agent on my account. For each one, give me a one-line read on whether it's working well, drifting, or unused. Then recommend specific actions: delete, merge, sharpen, or leave alone. Be blunt. Don't pad.

The Reviewer pulls your agent list, reads each Description and any attached context files, and produces the audit. Read it carefully — at minimum, you’ll learn things you didn’t know about your own fleet.

Step 5: Act on the obvious wins

From the audit, pick the moves that are obvious and easy:

• Delete agents that haven’t been used in 30+ days AND duplicate work covered elsewhere.
• Merge or split agents the audit flagged as overlapping.
• Sharpen Descriptions for the agents flagged as too vague — Tutorial 05 covers the editing.

Save the harder decisions for later. The audit is advisory, not prescriptive — your judgment overrides.

Step 6: (Optional) Schedule it monthly

Use Power Move 2 above. The Admin Agent walks you through wiring the Reviewer into a workflow that runs on the first Monday of every month and posts the audit to a Slack channel. Once that’s set up, the maintenance is on a clock and you stop having to remember.

What’s next

• → Tutorial 13: Run a workflow — needed for scheduling the monthly audit.
• → Tutorial 16: Connect Slack to your agents — needed for the audit to post somewhere your team will see.
• → Tutorial 25: Set up your Monday morning briefing — the same recurring-pattern idea applied to a different kind of report.

Reference

• The Agent Reviewer is just an agent — same governance, same edit flow as any custom agent.
• Disuse threshold: 30 days is a sensible default; adjust based on how seasonally your fleet works.
• Three drift patterns: overlap, scope creep, disuse.
• Action verbs: delete, merge, sharpen, leave alone.
• See also: Tutorial 07: Create an agent from scratch, Tutorial 13: Run a workflow, Tutorial 12: View your audit logs.