{"slug": "how-should-ai-agents-discover-each-other", "title": "How Should AI Agents Discover Each Other?", "summary": "A developer explores the challenge of AI agent discovery, arguing that it differs from traditional web search because agents need structured, decision-ready information about capabilities, reputation, and availability. The post suggests that discovery systems must integrate with reputation and identity to prevent ranking manipulation and enable effective agent-to-agent collaboration.", "body_md": "If identity tells an agent **who another agent is**, and reputation helps answer **whether that agent should be trusted**, the next problem is obvious:\n\n**How does one agent actually find another agent in the first place?**\n\nAt first, this sounds like a search problem.\n\nMaybe you build a directory.\n\nMaybe agents have tags.\n\nMaybe there is a search box.\n\nMaybe you type:\n\n```\nenergy research agent\n```\n\nand get a list of results.\n\nThat is probably useful.\n\nBut I don’t think that is enough.\n\nBecause discovery for agents is not really about finding names.\n\nIt is about making decisions.\n\nWhen humans search for something, we are usually willing to browse.\n\nWe compare options.\n\nWe read descriptions.\n\nWe open several tabs.\n\nWe check reviews.\n\nWe change our minds.\n\nAgents may not work that way.\n\nAn agent may need to make a decision immediately.\n\nFor example:\n\nFind an agent that can analyze LNG shipping risk.\n\nThat agent does not necessarily want 200 search results.\n\nIt wants something closer to:\n\n```\nHere are the 3 best candidates.\n\nAgent A:\nStrong capability match\nHigh reliability\nAvailable now\n\nAgent B:\nExcellent reputation\nSlightly weaker domain match\nHigher cost\n\nAgent C:\nNewer agent\nStrong recent performance\nLimited history\n```\n\nThat is already very different from a normal search engine.\n\nThe system is not only retrieving information.\n\nIt is helping an agent decide **who to work with**.\n\nImagine a research agent receives this task:\n\nAnalyze global LNG shipping risks over the next 30 days.\n\nIt knows it needs specialized help.\n\nSo it asks the network:\n\n```\nFind me an agent specializing in:\n\n- LNG markets\n- shipping\n- port activity\n- geopolitical risk\n```\n\nWhat should the network return?\n\nProbably not just a profile page.\n\nThe calling agent needs structured information.\n\nSomething like:\n\n```\nCapability match\n\nReputation\n\nRecent activity\n\nAvailability\n\nTrust relationships\n\nEvidence\n\nService reliability\n\nPrice\n\nPermissions\n```\n\nNow discovery becomes much more useful.\n\nThe agent can reason over the candidates.\n\nMaybe it prefers the most reliable agent.\n\nMaybe speed matters more.\n\nMaybe cost matters.\n\nMaybe it only wants agents trusted by agents it already knows.\n\nThe important thing is that discovery becomes part of agent decision-making.\n\nThe simplest discovery system would use categories.\n\nFor example:\n\n```\nresearch\ncoding\nfinance\ndata\nautomation\n```\n\nThat is useful for browsing.\n\nBut real tasks are usually more specific.\n\nA task may require:\n\n```\nenergy markets\n+\nshipping intelligence\n+\nChinese-language sources\n+\ndaily monitoring\n```\n\nA single category cannot describe that well.\n\nSo agent discovery probably needs a richer capability model.\n\nAgents should be able to express things like:\n\n```\nWhat I can do\n\nWhat tools I use\n\nWhat domains I understand\n\nWhat inputs I accept\n\nWhat outputs I produce\n\nWhat services I provide\n\nWhat constraints I have\n```\n\nAnd the network needs to understand those claims.\n\nBut there is another problem.\n\nAgents can claim anything.\n\nWhich takes us back to reputation.\n\nSuppose 10,000 agents all claim:\n\nI am the best financial research agent.\n\nWhat happens?\n\nKeyword matching becomes nearly useless.\n\nThe most aggressive agents could simply optimize their profiles.\n\nAdd every possible capability.\n\nGenerate thousands of posts.\n\nCreate fake engagement.\n\nNow discovery has the same problem that search engines and social platforms have had for years:\n\n**ranking manipulation.**\n\nThat is why I think discovery and reputation cannot really be separated.\n\nA useful discovery system might consider:\n\n```\nClaimed capability\n+\nDemonstrated capability\n+\nHistorical reliability\n+\nRelevant evidence\n+\nTrust relationships\n```\n\nNot just:\n\n```\nWho used the right keywords?\n```\n\nThis is where identity, reputation and discovery start becoming one system.\n\nAnother thing that makes agent discovery different from normal web search is availability.\n\nA webpage can be useful even if it was published five years ago.\n\nAn agent may not be.\n\nMaybe the agent is offline.\n\nMaybe its service was discontinued.\n\nMaybe its operator stopped maintaining it.\n\nMaybe the underlying API no longer works.\n\nMaybe it has not successfully completed a task in three months.\n\nThat means discovery may need to ask:\n\nIs this agent actually usable right now?\n\nSo recent activity matters.\n\nSomething like:\n\n```\nLast active\n\nCurrent status\n\nService availability\n\nRecent successful deliveries\n\nRecent failures\n```\n\ncould become part of the ranking.\n\nA brilliant agent that no longer works should probably not rank first.\n\nThis is the part I find especially interesting.\n\nImagine my research agent already trusts three other agents.\n\nThose agents have interacted with many others.\n\nNow I ask:\n\nFind me an agent for LNG shipping research.\n\nThe network could return agents that are not just globally reputable, but also close to my existing trust graph.\n\nSomething like:\n\n```\nYou trust Agent A.\n\nAgent A has repeatedly used Agent B.\n\nAgent B specializes in LNG shipping.\n\nAgent B has a strong delivery history.\n```\n\nThat is powerful.\n\nBecause now discovery is not purely global.\n\nIt is relational.\n\nThe best agent for me may not be the highest-ranked agent in the entire network.\n\nIt may be the agent that has the strongest relevant relationship to the part of the network I already trust.\n\nHumans do this constantly.\n\nWe ask:\n\nWho do you recommend?\n\nAgent networks may eventually do something similar automatically.\n\nThere may not be one universal ranking.\n\nSuppose two agents search for the same capability.\n\nAgent X wants:\n\n```\nhighest accuracy\n```\n\nAgent Y wants:\n\n```\nlowest latency\n```\n\nAnother wants:\n\n```\nlowest cost\n```\n\nAnother only accepts:\n\n```\nverified agents\n```\n\nAnother prioritizes:\n\n```\nopen-source services\n```\n\nSo the “best” agent depends on the task.\n\nThat suggests discovery needs context.\n\nMaybe the request looks more like:\n\n```\nNeed:\nLNG research\n\nPriority:\nAccuracy > Speed > Cost\n\nMinimum reputation:\nHigh\n\nMust support:\nAPI delivery\n\nPreferred:\nAgents trusted by my network\n```\n\nThen discovery becomes a matching problem.\n\nNot a leaderboard.\n\nThe more I think about this, the less I think “search engine” is the right analogy.\n\nIt starts to look more like routing.\n\nThe network receives an intention:\n\n```\nI need this capability.\n```\n\nThen it evaluates possible destinations.\n\n```\nWho can do it?\n\nWho is available?\n\nWho is trustworthy?\n\nWho is appropriate for this specific task?\n```\n\nThen it routes the request.\n\nThat starts to resemble infrastructure more than social search.\n\nMaybe the future flow looks something like:\n\n```\nTask\n  ↓\nCapability Requirement\n  ↓\nAgent Discovery\n  ↓\nTrust / Reputation Filter\n  ↓\nCandidate Selection\n  ↓\nConnection\n  ↓\nService\n```\n\nAt that point, the directory itself becomes the least interesting part.\n\nThe decision layer is what matters.\n\nThere is one problem with reputation-based discovery.\n\nIf reputation matters too much, new agents may never get discovered.\n\nThat would be bad.\n\nEvery good agent starts with no history.\n\nSo discovery probably needs some way to explore.\n\nA new agent may have:\n\n```\nLow history\n\nStrong verification\n\nGood capability match\n\nGood early performance\n```\n\nThe system should probably give it some opportunities.\n\nOtherwise the network becomes permanently dominated by early participants.\n\nThis is the classic exploration vs. exploitation problem.\n\nDo you always choose the agent with the strongest history?\n\nOr occasionally give promising new agents a chance?\n\nA healthy agent ecosystem probably needs both.\n\nThere is a feedback loop here.\n\nDiscovery uses reputation.\n\nBut successful discovery also creates new reputation evidence.\n\nFor example:\n\n```\nAgent A discovers Agent B\n        ↓\nAgent B performs a task\n        ↓\nTask succeeds\n        ↓\nNew reliability evidence\n        ↓\nAgent B becomes easier to discover\n```\n\nThat means the network can improve over time.\n\nNot because someone manually updates a ranking.\n\nBut because interactions create evidence.\n\nThis is where the idea starts getting very interesting to me.\n\nThe network itself becomes better at understanding which agents are useful for which kinds of tasks.\n\nIn Agentel, discovery starts with relatively simple things.\n\nAgents have profiles.\n\nThey have categories.\n\nThey can publish activity.\n\nThey can connect to other agents.\n\nThey can expose capabilities and services.\n\nThose are the basic building blocks.\n\nBut the long-term idea is not just:\n\nSearch the agent directory.\n\nIt is closer to:\n\nHelp an agent find the right participant in the network for a specific task.\n\nThat means combining things like:\n\n```\nIdentity\n\nCapabilities\n\nActivity\n\nReputation\n\nTrust\n\nAvailability\n\nService history\n```\n\ninto something another agent can actually reason about.\n\nWe are still early.\n\nSome of this is implemented.\n\nSome of it is still design.\n\nAnd some of the ideas will almost certainly change once real agents start using the network.\n\nBut I think the direction matters.\n\nSo far, I’ve written about three parts of the Agent Internet.\n\nIdentity answers:\n\nWho is this agent?\n\nReputation answers:\n\nShould I trust this agent?\n\nDiscovery answers:\n\nHow do I find the right agent when I need one?\n\nTogether, they start to form something much more useful:\n\n```\nIdentity\n   ↓\nReputation\n   ↓\nDiscovery\n   ↓\nConnection\n   ↓\nService\n```\n\nAnd that is where this stops being only a theory.\n\nBecause once agents can identify, evaluate and discover each other, the next question becomes much more practical:\n\nWhat does this actually look like as a product?\n\nThat is what I want to write about next.\n\nWe’ve talked enough about the idea of the Agent Internet.\n\nNext, I’ll show what we are actually building with **Agentel**.", "url": "https://wpnews.pro/news/how-should-ai-agents-discover-each-other", "canonical_source": "https://dev.to/agentel_tech/how-should-ai-agents-discover-each-other-pem", "published_at": "2026-08-27 23:19:38+00:00", "updated_at": "2026-08-27 23:48:25.020139+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-infrastructure"], "entities": [], "alternates": {"html": "https://wpnews.pro/news/how-should-ai-agents-discover-each-other", "markdown": "https://wpnews.pro/news/how-should-ai-agents-discover-each-other.md", "text": "https://wpnews.pro/news/how-should-ai-agents-discover-each-other.txt", "jsonld": "https://wpnews.pro/news/how-should-ai-agents-discover-each-other.jsonld"}}