DocSpring has published "Standard Support," a draft specification designed to solve a growing blind spot in the AI agent ecosystem: how autonomous agents report bugs and get help. Released on November 17, the spec provides a standardized convention for agents to interact with support systems, ensuring that errors encountered during API calls or tool usage don't just vanish into the void of a failed session. The initial implementation is already live on DocSpringβs MCP server as of November 18, marking a shift from one-way feedback loops to bidirectional communication between agents and human support teams.
The Three-Step Support Protocol
The spec defines three distinct tools to handle different levels of intervention, preventing agents from spamming support queues with trivial questions. The first, support_get_help, allows agents to query documentation instantly without user permission, acting as a self-service knowledge base. If that fails, support_send_feedback lets agents report friction or confusing errors asynchronously, with no expectation of a reply. Finally, support_open_ticket is reserved for blocking issues where a human answer is required; this action explicitly asks for user consent before submitting, ensuring the agent doesnβt act unilaterally on critical account issues.
Closing the Feedback Loop with Watch URLs
A key innovation in the spec is the watch_url attached to every ticket. Since most coding agents lack persistent email inboxes or webhook capabilities, the spec suggests agents run a background script to poll this URL for updates. This mechanism allows an agent to continue working on other tasks while waiting for a human response, then wake up to process the reply when it arrives. The URL itself is designed for privacy, revealing only that a change occurred, not the content of the message or the userβs account details, minimizing the risk of data leakage if the URL is exposed.
Security and Separation of Concerns
DocSpring deliberately separated the "Ask AI" chat feature for humans from the support_get_help tool for agents. While the human-facing chat can access account-specific context like recent invoices or template errors, the agent-facing help tool is restricted strictly to public documentation. This design choice prevents security leaks where a narrowly scoped agent token might accidentally expose sensitive billing data or template names through a help query. By keeping the agentβs help tool stateless and public-only, the spec ensures that support_get_help can run unattended without compromising user privacy or permission boundaries.
Key Takeaways
- The spec is currently at draft 0.4 and is available under a CC0 license, with code examples under MIT.
- Agents must pass
request_idsfrom API errors in tickets to help support teams trace logs quickly. - The protocol prevents prompt injection by treating all agent-submitted text as untrusted input.
- DocSpringβs CLI now supports these commands, allowing direct terminal-based ticket creation.
The Bottom Line
This is the missing infrastructure for agentic workflows; if your agent canβt file a ticket, itβs flying blind. Standard Support turns silent failures into actionable data, and I expect to see this pattern adopted by every major MCP server provider within the next quarter.