The Missing Piece in AI Agent Policy: can your bot make an account?
A new experimental proposal introduces /.well-known/agent-policy.json as a machine-readable manifest to standardize how websites communicate AI agent permissions, addressing a critical gap left by existing standards like robots.txt and llms.txt The key innovation is a tiered permission system rather than binary allow/deny, distinguishing between public exploration, declared automation via APIs, and conditional user-delegated signup with specific requirements The proposal defines "hard stops" tha
Analysis
TL;DR
- A new experimental proposal introduces /.well-known/agent-policy.json as a machine-readable manifest to standardize how websites communicate AI agent permissions, addressing a critical gap left by existing standards like robots.txt and llms.txt
- The key innovation is a tiered permission system rather than binary allow/deny, distinguishing between public exploration, declared automation via APIs, and conditional user-delegated signup with specific requirements
- The proposal defines "hard stops" that require agents to halt and request human intervention, including payment requirements, material liability terms, captcha challenges, and fake identity demands
- For agent builders, the standard provides a consistent framework to check permissions, implement safe fallbacks, and build transparent audit trails; for platform owners, it enables granular control over agent interactions while maintaining security
- The draft proposal from wkdomains.com is in experimental phase, open for community feedback from domain owners, bot builders, and standards organizations
Why It Matters
As AI agents evolve from passive information consumers to active delegated users performing account creation, form filling, and API interactions, the lack of a standardized policy mechanism creates friction for both website operators and agent developers. This proposal addresses a growing 2026 reality where server-side systems cannot distinguish legitimate agent-assisted actions from spam automation, potentially blocking useful agent functionality or exposing platforms to abuse.
Technical Details
- /.well-known/agent-policy.json: A new web standard proposal following the /.well-known/ convention, placing a machine-readable JSON manifest at the domain root that specifies granular permission tiers for AI agents
- Permission tier architecture: The manifest defines four operational modes—public exploration (read-only browsing), declared automation (via official APIs/sandboxes), conditional user-delegated signup (requiring verified human intent, transparent agent identity, traceable email patterns like user+agent-{domain}@example.org), and hard-stop conditions
- Hard stop triggers: Specific conditions that mandate agent termination and human takeover, including captcha/bot detection challenges, terms prohibiting automated accounts, fake identity requirements, payment methods, phone verification, public posting capabilities, material liability acceptance, and unknown terms impact
- Processing rules: Agents default to public exploration only when no manifest exists; explicit allows permit action within named limits with logging; routine checkboxes are allowed with disclosure; material terms trigger immediate stop
- Manifest structure: Includes metadata (schema version, domain, expiration date), account_creation section with default mode, available modes with conditions, terms acceptance logging, and hard stop definitions
Industry Insight
- Platform operators should monitor this proposal closely and consider implementing agent-policy.json manifests proactively to maintain control over agent interactions while avoiding blanket blocks that harm legitimate agent-assisted user experiences
- AI agent developers building user-delegated capabilities should adopt this standard early to ensure cross-platform compatibility and build trust through transparent audit trails rather than maintaining fragile custom logic per platform
- The distinction between routine checkboxes and material liability terms represents a crucial design pattern that could influence broader agent-policy standards, suggesting a model for handling graduated risk in automated actions across web services
Disclaimer: The above content is generated by AI and is for reference only.