Mozilla, the parent company of the Firefox browser, is building a new project called CQ. Led by engineer Peter Wilson, the initiative is positioned as "Stack Overflow for Agents." This open-source project aims to enable AI agents to discover and share collective knowledge.
Peter Wilson noted that "agents repeatedly encounter the same problems," leading to unnecessary work and token consumption when diagnosing and fixing these issues. By using CQ, agents will first query a shared knowledge base and contribute new solutions.
Currently, developers can use context files like agents.md, skill.md, or claude.md (for Anthropic's Claude Code) to guide agents. However, Wilson advocates for "a dynamic approach that earns trust over time, rather than relying on static instructions."
The CQ code is written in Python and is currently in the exploration and testing phase. It supports local installation and includes plugins for Claude Code and OpenCode. The project features a Docker container for running the API, an SQLite database, and an MCP (Model Context Protocol) server.
According to its architecture documentation, knowledge stored in CQ is divided into three levels: local, organizational, and "global shared." The "global shared" level refers to a publicly available instance of CQ. Knowledge units start with low confidence and are not shareable; however, as they are confirmed by other agents or humans, their confidence score gradually increases.
Will Mozilla host a public development instance of CQ? Wilson stated, "We have internally discussed distributed versus centralized sharing and the implications of each approach for the community."
"Personally, I think Mozilla.ai could try to help kickstart CQ by first providing an initial central platform for those wanting to explore shared public resources," he added. "That said, this needs to be done pragmatically. We want to validate user value as quickly as possible while also considering the trade-offs and risks associated with hosting a central service."
The CQ workflow, including agent and human-in-the-loop interactions.
It is evident that the project is vulnerable to malicious content and prompt injection attacks, where attackers could instruct agents to perform malicious tasks. Countermeasures against malicious content include anomaly detection, diversity requirements (confirmation from different sources), and Human-in-the-Loop (HITL) verification.
Developers view security as a primary concern for the CQ concept. "It sounds like a good idea until you imagine the security nightmare scenarios," one developer remarked.
Trusting AI agents to assign confidence scores to a knowledge base that is subsequently used by other AI agents could be problematic, as AI agents themselves can make mistakes or hallucinate. Human-in-the-Loop (HITL) technology can provide oversight for this process.
Regarding Stack Overflow, Wilson used the term "matriphagy"—where offspring consume the parent—to describe its decline. He wrote, "Large Language Models (LLMs) have committed matriphagy against Stack Overflow through agents. Now, agents need to build their own Stack Overflow."
The number of questions on the Stack Overflow community has dropped sharply. Currently, the company owns an MCP server to store its content and is positioning its private product, Stack Internal, as a way to provide knowledge to AI.
Why is Mozilla doing this? According to its State of Mozilla report, the non-profit organization is "reinventing Mozilla to succeed in AI just as it did in the web." Mozilla.ai, part of the Mozilla Foundation, includes projects like Octonous for managing AI agents and any-llm, which provides a unified interface for multiple LLM providers.
Mozilla also operates the widely popular MDN (Mozilla Developer Network) documentation site, a comprehensive reference platform providing documentation for JavaScript, CSS, and Web APIs.
So far, reassuringly, it has not utilized artificial intelligence.
cq: Stack Overflow for Agents: https://blog.mozilla.ai/cq-stack-overflow-for-agents/
GitHub - mozilla-ai/cq: https://github.com/mozilla-ai/cq
Author: Chang Chang