Skip to content

Latest commit

 

History

History
338 lines (239 loc) · 10.3 KB

File metadata and controls

338 lines (239 loc) · 10.3 KB
title Agent Approval Patterns
slug agent-approval-patterns
category ai
depth 2
audit_level
2
3
4
last_reviewed 2026-04-21
sources
NIST AI RMF 1.0 — https://doi.org/10.6028/NIST.AI.100-1 (2023-01)
NIST AI 600-1 Generative AI Profile — https://doi.org/10.6028/NIST.AI.600-1 (2024-07)
OWASP Top 10 for LLM Applications v2025 — https://genai.owasp.org/llm-top-10
Anthropic — Building Effective Agents — https://www.anthropic.com/research/building-effective-agents (2024)
OpenAI Preparedness Framework — https://cdn.openai.com/openai-preparedness-framework-beta.pdf (2023-12)
triggers_strong
agent approval
human in the loop
agent autonomy
approval workflow
tool use approval
triggers_weak
approval pattern
agent permissions
HITL
related
llm-agent-security
ai-tool-profiles
prompt-and-tool-evidence-handling
browser-computer-use-security

Agent Approval Patterns

Last reviewed: 2026-04-18 | Next review: 2026-10-18 | Priority: Essential | Audit Level: 2-4 | Automation: Partial (risk tiering, policy enforcement, and some pre-flight checks automatable; business-context review, change judgment, and exception approval manual)

Use this guide when deciding what an AI agent may do without review, what requires explicit approval, and what should never be delegated directly.

The goal is not to slow systems down with approval theater. The goal is to place review exactly where blast radius and reversibility demand it.


1. Core rule

Approval should be driven by:

  • capability,
  • trust boundary crossed,
  • data sensitivity,
  • blast radius,
  • reversibility,
  • operator confidence.

A simple chat task and a production deployment are not in the same approval class, even if both are performed by the same model.


2. Five approval tiers

Tier Typical capability Approval expectation
T1 Reviewer read-only analysis, summarize, classify, suggest no approval beyond normal access policy
T2 Editor draft code or text, propose changes, write local artifacts not yet applied human review before commit / publish
T3 Integrator modify repo, update configs, run bounded commands, fetch docs approval before merge or before executing impactful changes
T4 Operator touch cloud resources, secrets, tickets, CI, or user-facing systems explicit task-level approval and scoped credentials
T5 Admin production mutations, identity changes, financial or legal impact, destructive actions strong human approval, dual control where justified

The same agent may operate at different tiers depending on the tool and context. Tier is about action, not branding.


3. Actions that should almost always require approval

Require explicit human approval for actions such as:

  • production deploys,
  • granting or modifying admin access,
  • secret rotation that can break systems,
  • creating or approving firewall / IAM exceptions,
  • deleting data or infrastructure,
  • sending externally visible communications on behalf of the company,
  • changing billing, payroll, or regulated records,
  • bulk export of sensitive data,
  • browser-based actions in third-party admin consoles.

If the action can create a material incident in one step, approval should not be implicit.


4. Approval dimensions

4.1 Capability

What can the agent do?

  • read files,
  • write files,
  • execute commands,
  • access the network,
  • browse the web,
  • call internal APIs,
  • touch production systems,
  • handle secrets or sensitive data.

4.2 Context

Where is it acting?

  • local sandbox,
  • source repository,
  • CI environment,
  • ticketing system,
  • staging,
  • production,
  • third-party SaaS admin panel.

4.3 Reversibility

Can the action be rolled back safely?

  • drafting text: easy,
  • code edit before merge: moderate,
  • migration against production data: hard,
  • deleting accounts or records: often hard or impossible.

4.4 Observability

Can a reviewer see what happened?

  • input and output preserved,
  • tool calls logged,
  • approvals recorded,
  • artifacts retained,
  • rollback path known.

Poor observability should increase approval requirements.


5. Default approval matrix

Action class Example Suggested tier
read-only reasoning review docs, classify risk, summarize logs T1
draft-only changes propose code patch, draft policy text T2
repo write in feature branch update tests, refactor code, regenerate docs T2-T3
network fetch / external retrieval fetch package docs or vendor pages T2-T3 depending on trust
CI or workflow changes edit pipeline config, permissions, deploy logic T3-T4
browser admin actions click through cloud console or SaaS admin T4-T5
production state changes deploy, rotate critical secret, delete resource T5
bulk data export / deletion DSAR export, incident data pull, cleanup jobs T4-T5

Adjust upward if the environment is regulated, customer-facing, or poorly observable.


6. Approval patterns that work well

6.1 Propose-then-approve

Pattern:

  1. agent analyzes and prepares a plan,
  2. human reviews the proposed action set,
  3. agent executes only the approved subset,
  4. output and evidence are retained.

Use for:

  • code changes,
  • policy updates,
  • infrastructure changes with bounded scope.

6.2 Dry-run before execution

Pattern:

  1. agent generates diff, command plan, or simulated API call,
  2. human reviews expected blast radius,
  3. agent executes with same parameters,
  4. post-action verification runs.

Use for:

  • Terraform plan/apply,
  • batch user changes,
  • CI config changes,
  • admin operations with preview support.

6.3 Two-person approval for high-impact actions

Pattern:

  • one approver confirms business intent,
  • one approver confirms technical safety.

Use for:

  • production deletions,
  • privileged IAM changes,
  • irreversible migrations,
  • emergency changes under pressure.

6.4 Bounded autonomy window

Pattern:

  • grant temporary scope for a narrow task,
  • restrict tools, targets, and duration,
  • require re-approval when scope changes.

Use for:

  • incident triage,
  • repo-wide cleanup,
  • time-boxed research or migration work.

7. Approval anti-patterns

Avoid these patterns:

  • blanket approval: "the agent can do anything in this environment"
  • sticky approval: one old approval silently covers future unrelated actions
  • approval by obscurity: reviewer cannot see the diff, command, or target
  • approval after execution: human review exists only on paper
  • identity mismatch: unclear who approved versus who executed
  • shared privileged tokens: approval cannot be tied to a scoped credential

A bad approval flow creates false confidence instead of safety.


8. Role design for humans in the loop

Define who can approve what.

Human role Typical authority
Task requester confirms business intent and desired outcome
Code reviewer approves repo changes before merge
System owner approves changes to owned service or environment
Security reviewer approves high-risk control exceptions or security-sensitive actions
Incident commander authorizes urgent containment and recovery actions

Do not let "somebody in chat said yes" be your approval model for high-impact actions.


9. Approval evidence to retain

For non-trivial actions, retain:

  • request or task ID,
  • actor and environment,
  • proposed action summary,
  • exact diff / command / target where possible,
  • approver identity,
  • approval timestamp,
  • execution result,
  • rollback or verification outcome.

This matters for incident response and for proving the approval was real.


10. Special handling for browser and computer use

Browser-capable agents need stricter approval because they can interact with dynamic, ambiguous interfaces.

Require stronger review when the agent can:

  • open third-party admin portals,
  • upload files,
  • approve payments or transactions,
  • change IAM or billing settings,
  • click through security prompts,
  • operate in production consoles.

Prefer:

  • pre-defined targets,
  • read-only browsing by default,
  • screenshot or action logs,
  • human confirmation before final submission actions.

11. Sensitive data and regulated workflows

Increase approval requirements when the task involves:

  • HR or employee data,
  • legal or investigation material,
  • customer exports,
  • financial records,
  • credentials or secrets,
  • regulated environments with mandatory evidence trails.

If the agent has broad data visibility, use smaller approval scopes and stricter evidence handling.


12. First 30 minutes of designing an approval policy

  1. List tools and environments the agent can reach.
  2. Classify actions into read, write, execute, network, admin, and destructive.
  3. Map those actions to trust boundaries and data sensitivity.
  4. Decide which actions are never automatic.
  5. Define approver roles and escalation path.
  6. Define what evidence must be logged.
  7. Test one safe task and one high-risk task against the policy.

If the policy cannot explain those seven points, it is too vague.


13. Common red flags

Red flag Why it matters
one approval covers all future actions approval drift
agents can use prod credentials without scoped gating concentrated blast radius
browser actions submit forms without confirmation easy irreversible mistakes
approval logs do not include exact target or diff weak accountability
human reviewers cannot inspect the proposed action clearly rubber-stamping risk
emergency paths bypass all logging incident-within-incident risk

14. Minimum checklist

Check Expected
Agent capabilities are classified by tier and environment Yes
High-impact or irreversible actions require explicit approval Yes
Approval is tied to named humans and scoped credentials Yes
Proposed actions are reviewable before execution Yes
Browser and production-console actions have stricter controls Yes
Sensitive-data workflows use tighter approval scopes Yes
Execution and approval evidence are retained for audit/IR Yes
Emergency paths are logged and reviewed after use Yes

15. Related references

  • llm-agent-security.md
  • ai-tool-profiles.md
  • prompt-and-tool-evidence-handling.md
  • browser-computer-use-security.md