AI Agent Allowlist
Home Page-Types Database Agent Guardrails 2026 Incidents API Docs Pricing
Resources
Use Cases (15) Industries & Buyers (12) Learn: Core Concepts (12) Implementation Guides (15) Comparisons (8) Schema & Data Reference (6) FAQ Glossary
Why It Matters
2026 Agent Incidents Category Targeting Database Refreshes Contact Customer Login
Download Free Sample
six real files, validated, free to download

AI Agent Rules File Examples

Six ready-made rules files for common agent roles. Each one says which pages the agent may open, what happens to unknown sites, and which narrow exceptions it has.

Every example is checked with the free Agent Egress Guard and uses page types from an AI agent allow list.

6Example files
4Agency tiers used
1Purchase exception
JSONPlain, portable format
Before you copy

How to read these examples

A rules file is a promise about what an agent may do on the web. Copy the structure, then change the details to match your agent's real purpose.

Each example below shows the file on the left and the reasoning on the right. Download links give you the exact file, ready to validate.

Purpose first

Each allow list follows the agent's job. A support agent has no reason to open careers pages.

Actions stay closed

No example allows signup, checkout, upload or posting as a general page type.

Exceptions are narrow

The one purchase permission names a page type, a domain and an end date.

Every file has an owner

And a review date. Validation warns when either is missing.

Example 1

Market research agent

Compares vendors, reads prices and product pages, writes a summary. Tier 3: it plans its own steps, a person approves anything risky.

{ "agent": "market-research", "owner": "head-of-strategy", "tier": 3, "review_by": "2026-12-31", "web": { "allow_page_types": ["pricing", "documentation", "blog", "press", "case_studies", "about", "product"], "unclassified": "deny", "on_deny": "ask_owner", "exceptions": [] } }

Why each line

  • Seven read types cover everything a comparison needs.
  • Unknown sites are denied, because research links often lead to file shares and trackers.
  • Denials go to the owner, so a blocked "see full pricing" signup becomes a decision, not a silent failure.
  • No exceptions: research never needs to buy or post.

Good fit for competitive analysis, analyst research and vendor comparisons.

market-research.json
Example 2

Customer support answer agent

Answers customer questions using vendors' help pages and status pages. Tier 2: it follows a fixed workflow.

{ "agent": "support-answers", "owner": "support-lead", "tier": 2, "review_by": "2027-03-31", "web": { "allow_page_types": ["help_center", "status", "documentation", "contact"], "unclassified": "deny", "on_deny": "block", "exceptions": [] } }

Why each line

  • Four read types match the workflow exactly.
  • Block, not ask: a support workflow runs at volume, and nobody can approve each request.
  • A longer review date suits a narrow, read-only agent.
  • Contact pages are allowed so the agent can point customers to the right team, not so it can submit forms.

Good fit for help desks that answer questions about third-party products.

support-answers.json
Example 3

Procurement agent with one purchase exception

Researches suppliers and places small orders with one approved supplier. Tier 3, with owner approval per order.

{ "agent": "procurement", "owner": "procurement-lead", "tier": 3, "review_by": "2026-12-31", "web": { "allow_page_types": ["pricing", "legal", "security", "about", "documentation", "case_studies"], "unclassified": "deny", "on_deny": "ask_owner", "exceptions": [{ "page_type": "checkout", "domain": "approved-supplier.example", "expires": "2027-03-31", "requires_approval": true }] } }

Why each line

  • Legal and security pages let the agent check contract and data terms.
  • The exception opens checkout on one domain only. Every other checkout stays denied.
  • requires_approval means each order waits for a person.
  • The end date forces a renewal decision in March.

Pair it with a spend limit in your purchasing system, so a mistake has a ceiling.

procurement.json
Example 4

Coding assistant

Reads documentation while writing code in a sandbox. Tier 4: it acts on its own for long stretches.

{ "agent": "coding-assistant", "owner": "platform-engineering", "tier": 4, "review_by": "2026-12-31", "web": { "allow_page_types": ["documentation", "status", "integrations"], "allow_domains": ["docs.internal.example"], "unclassified": "allow_reads", "on_deny": "block", "exceptions": [] } }

Why each line

  • Documentation, status and integration pages cover most coding research.
  • Unknown reads are allowed on purpose: code answers live on many small sites. The validator warns about this, and the warning is accepted in review.
  • Unknown writes are still denied by the baseline, so uploads and posts cannot happen.
  • Block, never ask: nobody watches a tier 4 agent in real time.

Stricter teams switch unclassified to deny and list their approved documentation domains.

coding-assistant.json
Example 5

Sales research agent

Builds account briefs from company websites. Tier 3.

{ "agent": "sales-research", "owner": "sales-ops", "tier": 3, "review_by": "2026-12-31", "web": { "allow_page_types": ["about", "leadership", "careers", "case_studies", "press", "blog", "partners"], "unclassified": "deny", "on_deny": "ask_owner", "exceptions": [] } }

Why each line

  • Leadership, careers and press pages describe a company's priorities and growth.
  • Contact pages are left out on purpose, so the agent cannot fill in sales forms under your brand.
  • Comment and signup pages stay denied, which protects your reputation.
  • Owner routing catches the rare case where a form really is needed.

Good fit for account research, lead enrichment and meeting preparation.

sales-research.json
Example 6

Compliance monitoring agent

Watches vendors' legal, security and status pages for changes. Tier 2.

{ "agent": "compliance-monitor", "owner": "compliance-lead", "tier": 2, "review_by": "2027-03-31", "web": { "allow_page_types": ["legal", "security", "status", "about"], "unclassified": "deny", "on_deny": "block", "exceptions": [] } }

Why each line

  • Four read types are all a change monitor needs.
  • Verified legal and security URLs avoid guessing where a vendor keeps its terms.
  • Block on deny: a monitor should never do anything but read.
  • An audit trail of visits doubles as evidence of vendor monitoring.

Good fit for vendor risk teams that track terms and security page changes.

compliance-monitor.json
Tests

Golden tests for each example

Keep these next to each file and run them on every change. The expected verdicts assume the page-type database is connected.

FileRequestExpected
market-researchGET a vendor's verified pricing pageallow
market-researchGET the same vendor's free-trial signupapproval_required
support-answersGET a vendor's status pageallow
support-answersGET a vendor's careers pagedeny
procurementGET checkout on approved-supplier.exampleapproval_required
procurementGET checkout on any other shopapproval_required, then refused
coding-assistantGET an unknown documentation siteallow
coding-assistantPOST to a package upload URLdeny
sales-researchPOST to a contact formapproval_required
compliance-monitorGET a cloud metadata addressdeny

The procurement row shows why owner approval matters: the request reaches a person, who says no, and the refusal is logged.

Add at least one test per page type the agent is allowed, and one per action type it could plausibly meet.

Adapting by sector

Changes regulated teams usually make

The examples are a neutral starting point. Regulated sectors usually tighten them in these ways.

Financial services

Move every agent to unclassified: deny, including coding assistants.

Add deny_domains for data brokers and file-sharing sites.

Healthcare

Keep uploads denied everywhere, with no exceptions.

Shorten review dates to 60 days for agents that touch patient data.

Public sector

Use tier 2 wherever possible, with block rather than ask.

License the page-type data on-premise so no lookup leaves the network.

Legal teams

Allow legal and security pages widely, deny comment and post everywhere.

Record every visit for matter files.

Before adopting

Checklist for turning an example into your policy

1

Match the purpose

Write the agent's purpose in one sentence, then remove any read type it does not serve.

2

Name real people

Replace the owner with one named person from your agent register.

3

Check the tier

Look at the agent's tools. Browsing plus writing tools usually means tier 3 or 4.

4

Replace example domains

Every domain ending in .example must become a real one, or be removed.

5

Run validate and the tests

Fix every error. Reply to every warning in the pull request.

Side by side

The six files compared

One table makes the differences between the six roles easy to see and discuss in a review.

AgentTierRead typesUnknown sitesOn denyExceptions
market-research37DenyAsk ownerNone
support-answers24DenyBlockNone
procurement36DenyAsk ownerCheckout, one supplier
coding-assistant43Allow readsBlockNone
sales-research37DenyAsk ownerNone
compliance-monitor24DenyBlockNone

Notice the pattern. Tier drives on_deny, purpose drives the read types, and only one file needs an action page at all.

That is typical in real deployments. Most agents only need to read, and the few that must act can do so through one narrow, dated exception.

If your own files need many exceptions, the agent is probably doing too many jobs. Split it into two agents with two files.

Using them

From download to enforcement

Three commands take a file from download to a working check.

# 1. install the free enforcer pip install https://www.aiagentallowlist.com/egress-guard/agent_egress_guard-0.2.0-py3-none-any.whl # 2. check the file agent-egress-guard policy validate procurement.json # 3. try a request agent-egress-guard policy check procurement.json https://approved-supplier.example/checkout APPROVAL_REQUIRED rule: exception

Rename the agent and owner

Use the names from your agent register, so logs and audits match.

Adjust the read types

Remove any type the agent does not need. Fewer is safer.

Set your own dates

Review dates and exception end dates should reflect your calendar, not ours.

Add golden tests

Five to ten URLs per file with the verdicts you expect.

Page types

The vocabulary the files use

Read types can go on an allow list. Action types can only be opened through an exception.

aboutblogcareerscase_studiescommunitycontactdocumentationeventshelp_centerintegrationsleadershiplegalpartnerspresspricingproductsecuritysitemapstatus
loginsignuppassword_resetcartcheckoutsubscribeuploadpost_createcomment

Full definitions: page types explained.

Mistakes

What goes wrong when teams write their own

These six mistakes appear again and again in first drafts. Each one widens what the agent can do without anyone deciding it should.

Allowing every read type

Convenient at first, but the agent then wanders into pages its task never needed.

Checkout as a page type

That opens checkout everywhere. Use an exception for one domain instead.

No end date on exceptions

Temporary permissions quietly become permanent.

Ask-owner on tier 4

Nobody is there to answer, so the agent just stalls.

Shared files for different agents

The shared file ends up as open as the most demanding agent.

Owner set to a team alias

An alias cannot be accountable. Name one person.

Versioning

Keeping rules files healthy over time

A file written once and never touched drifts away from what the agent really does. Small, regular habits keep it accurate.

One repository

All agent files in one folder, with the agent owners as required reviewers.

Review reminders

A weekly job lists files whose review date is within 14 days.

Expiry sweeps

A monthly job removes expired exceptions in a pull request, so the file matches reality.

Release tags

Tag the repository when files change, and record the tag in each verdict log.

Diff in audits

Auditors can compare two tags and see every permission granted between them.

Delete on retirement

Removing the file removes the permissions. The history stays in version control.

Terms

Words used in the files

allow_page_types

The read page types an agent may open. Empty means all read types.

unclassified

What happens on pages no layer can identify: deny, or allow reads.

on_deny

Block a denied request, or route it to the owner for a decision.

exception

A narrow, dated permission for one page type on one domain.

requires_approval

Each use of the exception waits for the owner to approve.

tier

How far the agent may act alone, from 1 (answers only) to 4 (acts on its own).

Tested against the 2026 agent incidents

  • None of the six files allows uploads, posts, wiki edits or plugin installs.
  • Those were the steps in the documented 2026 incidents.
  • In our replay, page data plus egress rules would have stopped almost all of them.
See the incident-by-incident prevention analysis The dataset upload case

The honest fine print — the same two assumptions we publish, plus two operational ones

  1. The policy engine must see every request — an agent with raw socket access or a second network path bypasses everything; enforcement belongs at the egress proxy/network layer, not only in an SDK hook.
  2. Default-deny must be on. In flag-only mode these become alerts within minutes rather than prevention — still a large improvement on a timeline measured in weeks (the DseWiki edits ran from late May to late June 2026, per the researchers), but not a block.
  3. For full URL+method matching on HTTPS you need to be the proxy or in-process hook — SNI alone shows only the host, which still catches the entire host-list layer.
  4. Policy can’t read intent inside a legitimately allowed action: an agent whose job is publishing packages keeps registry access. In our replay of the 2026 incidents, no crossing fits any plausible allowlist for the agents’ documented tasks.
Related

Keep reading

FAQ

Rules file questions

What is an AI agent rules file?
A small file that states what one agent may do on the web: which page types it may open, which domains are always allowed or denied, and which narrow exceptions apply.
Can I use these files as they are?
They are valid and enforceable, but change the agent name, owner, dates and domains to match your own setup before relying on them.
Why does the coding assistant allow unknown reads?
Answers to coding questions live on many small sites. The trade-off is accepted on purpose, and unknown writes remain denied by the baseline.
What enforces these files?
The free Agent Egress Guard from version 0.2.0. The format is plain JSON, so other gateways and proxies can apply it too.
How does the enforcer know what page type a URL is?
It looks it up in page-type data. An AI agent allow list covers 28 page types on 40M+ domains, from $99 a month. See pricing.
Do the example domains work?
No. Domains ending in .example are placeholders reserved for documentation. Replace them with real domains before using a file.
Should every agent have an exception list?
No. Most agents need none. An empty exceptions list is a sign of a well-scoped agent.
Can one file cover several agents?
It can, but it should not. Separate files keep each agent's permissions as narrow as its own purpose.

Download a file, change four lines, enforce it today

Free examples, free enforcer, page-type data when you need full coverage.

Get an example