How to Set Up OpenClaw Autonomy Tiers for Your Business

Why Autonomy Tiers Matter

The single most common mistake operators make when configuring an OpenClaw agent is treating autonomy as binary: either the agent can do something or it can't.

Real operations don't work that way. Some tasks are low-risk and high-frequency — you want the agent running them without interrupting you every five minutes. Others are medium-stakes — you want a heads-up but not a full pause. And some should always come back to you before anything happens.

That's the three-tier autonomy model. When it's configured correctly, your agent moves fast on the things you trust, slows down on the things you want visibility into, and stops completely on the things that need your judgment.

What the Three Tiers Look Like

Here's the framework from a production AGENTS.md:

```

Autonomy Tiers

### Tier 1 — Full Autonomy (No Approval Needed)

- Draft and save internal documents

- Run scheduled reports and status checks

- Read files, search repos, pull data

- Update memory files and session logs

- Generate content drafts for review queue

### Tier 2 — Notify Then Execute (Proceed Unless Stopped)

- Send pre-approved templates to known contacts

- Create calendar events or task items

- Post to internal Slack channels

- Trigger automated sequences already in the approved playbook

### Tier 3 — Explicit Approval Required

- Send any first-contact outreach

- Publish anything externally (posts, pages, emails to new lists)

- Spend money or initiate billing actions

- Delete, archive, or move files permanently

- Modify AGENTS.md, SOUL.md, or core config files

```

The exact contents of each tier will vary by business. A solo operator running a content agency has different Tier 1 defaults than a freelancer managing client contracts. The structure is fixed; the contents are yours to set.

How to Decide What Goes Where

Use three questions to place any action in a tier:

1. What is the blast radius if this goes wrong?

A draft saved to the wrong folder is recoverable in seconds. An email sent to the wrong client is not. The size of the cleanup determines the tier.

2. How often does this happen?

High-frequency tasks in Tier 3 will kill your productivity — you'll spend half your day approving routine actions. If something happens more than three times a day, it probably belongs in Tier 1 or Tier 2.

3. Can I undo this?

Irreversible actions belong in Tier 3 by default. Deletions, external sends, billing actions, config changes — if you can't easily reverse it, require approval.

Common Mistakes

Putting too much in Tier 1. This usually happens when operators are excited about automation and want to see everything running. The problem surfaces the first time the agent does something sensible but wrong — publishes a draft that wasn't ready, sends a follow-up at the wrong time, updates a file you were still editing. Start conservative and loosen the tiers as you build trust.

Putting too much in Tier 3. The opposite failure mode. Operators who've been burned by early automation often lock everything down. Now the agent is basically a chatbot again — it can't do anything without you approving it first. If you're approving more than five things a day, your Tier 3 is too broad.

Not defining escalation behavior. What does the agent do when it hits a Tier 3 situation at 2 AM when you're not available? Your AGENTS.md should include an explicit fallback: "If a Tier 3 action is triggered outside operating hours, log it to PENDING.md and notify via Telegram at next check-in."

A Real Tier 2 Example

Here's how a well-configured Tier 2 entry looks:

```

### Tier 2 — Follow-Up Sequences

If a lead responds to initial outreach and the response matches a known category

(interested, not now, wrong person), send the corresponding pre-approved template

from /templates/. Notify operator via Telegram with: lead name, template sent,

timestamp. Operator has 15 minutes to intervene. Log outcome to /memory/crm-log.md.

```

Notice what's happening here: the agent acts, but you get a window to intervene. The notification comes before you'd realistically see it, so you're not a bottleneck — but you're still in the loop.

Start With a Minimum Viable AGENTS.md

If you're setting this up from scratch, here's the fastest path:

1. List the ten most frequent tasks your agent will handle

2. Sort them by blast radius (low → high)

3. Put the bottom third in Tier 1, middle third in Tier 2, top third in Tier 3

4. Run it for a week and adjust based on what actually needed your attention

Autonomy tiers aren't a one-time configuration. They evolve as you build trust with the system, as your business changes, and as you discover actions you hadn't anticipated.

The goal is a system where you're only ever making decisions that actually require your judgment. Everything else should run without you.