[{"data":1,"prerenderedAt":175},["ShallowReactive",2],{"\u002Fblog\u002F2025-04-15-giving-a-community-agent-tools":3},{"id":4,"title":5,"archiveType":6,"body":7,"category":159,"contextDate":160,"cover":161,"coverAlt":162,"date":163,"description":164,"extension":165,"featured":166,"historicalDate":163,"historicalLabel":160,"meta":167,"navigation":168,"path":169,"publishedDate":160,"seo":170,"stem":171,"thumbnail":160,"workPeriod":172,"writtenDate":173,"__hash__":174},"blogs\u002Fblog\u002F2025-04-15-giving-a-community-agent-tools.md","Giving a community agent tools, and the ones it should never get","retrospective",{"type":8,"value":9,"toc":150},"minimark",[10,14,19,48,51,55,58,61,64,71,75,101,105,137,140,144,147],[11,12,13],"p",{},"In December 2024 the Playermon team previewed Razortooth AI, an in-house agent for the community's Telegram group. It took part in the chat and helped monitor the group, and I oversaw the work. This post is not about how it was built. It is about a question that has become more pressing in the last few weeks: what happens when giving an agent a new ability stops being an engineering project and becomes a line of configuration?",[15,16,18],"h2",{"id":17},"what-changed-in-the-last-few-months","What changed in the last few months",[20,21,22,30,36,42],"ul",{},[23,24,25,29],"li",{},[26,27,28],"strong",{},"The Model Context Protocol."," Anthropic published MCP as an open standard in November 2024. A server describes its tools, each with a name, a description and an input schema, and a client lets a model discover and call them.",[23,31,32,35],{},[26,33,34],{},"Wider adoption."," On 26 March OpenAI said it would support MCP, starting with its Agents SDK.",[23,37,38,41],{},[26,39,40],{},"A new revision of the specification,"," dated 26 March, added an authorisation framework based on OAuth 2.1 and tool annotations that describe whether a tool only reads or can destroy something.",[23,43,44,47],{},[26,45,46],{},"Agents talking to agents."," On 9 April Google announced the Agent2Agent protocol, A2A, so that agents from different vendors can hand tasks to each other. MCP connects an agent to tools; A2A connects it to other agents.",[11,49,50],{},"The practical effect is that the cost of adding a capability to an agent is falling quickly. That is exactly the moment to slow down and ask which capabilities it should have.",[15,52,54],{"id":53},"in-a-public-chat-every-message-is-input","In a public chat, every message is input",[11,56,57],{},"An agent in a community group is not like a coding assistant on my own laptop. Its users are many, mostly anonymous, and some of them are actively trying to make it misbehave. Every message in the group is text the model reads.",[11,59,60],{},"Without tools, a successful manipulation produces a bad reply. A moderator can delete it and apologise. With tools, a successful manipulation produces an action.",[11,62,63],{},"The risk is not only in the chat. On 1 April Invariant Labs described what they called tool poisoning: instructions hidden in a tool's description, which the model reads and the person using it usually never sees. The MCP specification is direct about this kind of risk. It says clients must treat tool annotations as untrusted unless they come from trusted servers, and that there should always be a human in the loop who can deny a tool call.",[11,65,66,67,70],{},"So the rule I would use for any community agent is simple. ",[26,68,69],{},"A tool given to an agent in a public group is a tool given to every member of that group."," If I would not put a button for it in front of an anonymous member, the agent should not have it either.",[15,72,74],{"id":73},"tools-i-would-give-it","Tools I would give it",[20,76,77,83,89,95],{},[23,78,79,82],{},[26,80,81],{},"Read-only search over published material:"," guides, patch notes, the FAQ and past announcements. Most questions in a game community are already answered somewhere, and pointing to the source is better than paraphrasing it.",[23,84,85,88],{},[26,86,87],{},"Read-only public chain lookups,"," such as whether a transaction a player has posted has confirmed. The data is public anyway; the agent only saves someone a trip to the block explorer.",[23,90,91,94],{},[26,92,93],{},"Flagging a message for a human moderator,"," with the reason attached.",[23,96,97,100],{},[26,98,99],{},"Short, reversible moderation,"," such as muting an account that is posting links in bulk, written to a log that a person reviews.",[15,102,104],{"id":103},"tools-it-should-never-get","Tools it should never get",[20,106,107,113,119,125,131],{},[23,108,109,112],{},[26,110,111],{},"Anything that signs."," No transfers, claims, airdrops or approvals, and no key material anywhere the agent can reach.",[23,114,115,118],{},[26,116,117],{},"Choosing winners or paying rewards."," Contests in a token community attract people who try to game them. A person should own those decisions and be accountable for them.",[23,120,121,124],{},[26,122,123],{},"Changing anyone's account,"," whether an email address, a linked wallet or a username.",[23,126,127,130],{},[26,128,129],{},"Posting in the official announcement channel."," A false announcement from an official account is a ready-made phishing campaign.",[23,132,133,136],{},[26,134,135],{},"Permanent bans,"," or anything else a member cannot appeal to a person.",[11,138,139],{},"And one rule that is not a tool at all: the agent must never ask anyone for a seed phrase or private key, and it should say plainly that the team never will.",[15,141,143],{"id":142},"what-the-protocols-do-not-decide","What the protocols do not decide",[11,145,146],{},"Annotations are hints from whoever wrote the server. Authorisation tells you who connected, not whether a particular call is wise. A2A lets one agent ask another for help, which adds another source of instructions to check. None of that is a criticism of the protocols. They standardise how tools are connected. They cannot decide which tools belong in a given room.",[11,148,149],{},"That decision stays with whoever is responsible for the agent. For each new tool I would ask two questions. What is the worst thing a hostile member could get it to do? Can a person undo it afterwards? If the answers are \"something serious\" and \"no\", the tool stays out, however easy it has become to plug in.",{"title":151,"searchDepth":152,"depth":152,"links":153},"",2,[154,155,156,157,158],{"id":17,"depth":152,"text":18},{"id":53,"depth":152,"text":54},{"id":73,"depth":152,"text":74},{"id":103,"depth":152,"text":104},{"id":142,"depth":152,"text":143},"AI",null,"\u002Fimg\u002Fblog-covers\u002F2025-04-15-giving-a-community-agent-tools.webp","A small robot beside an open toolbox, a key and a wallet locked in a separate glass case.","2025-04-15","Tool protocols such as MCP make it cheap to connect an agent to new capabilities. For a bot in a public chat, the list of tools is the list of permissions.","md",false,{},true,"\u002Fblog\u002F2025-04-15-giving-a-community-agent-tools",{"title":5,"description":164},"blog\u002F2025-04-15-giving-a-community-agent-tools","2025","2026-11-07","nm5jvSXihs4qqx0piSozMTnHt26vy9BMkU8ZZhnu_qI",1794838285269]