Talk:AI agent

Proposed Addition: Functional Framework Section

The current page provides a good basic definition, we noticed a gap in describing the generalized architecture that enables an agent to perceive, reason, and act. we plan to add a new section titled Functional Framework. We intend to break down an AI agent into its core components:

Perception: how agents receive information and sense their environment

Memory: Describe short-term memory and long-term memory

Planning: Break down complex problems into smaller subtasks

Action: How agents execute tasks

Environment: Defining the environment where the agent functions Lian0190 (talk) 19:52, 29 March 2026 (UTC)

Note that Wikipedia doesn't publish original research. Start with reliable, independent sources which are specifically about AI agents. Do not cite pre-prints (see WP:ARXIV). Grayfell (talk) 08:47, 4 April 2026 (UTC)

Pre-prints are not blanket unallowed, you can cite them under the same conditions as other acceptable SPS. Jacob Rampino (talk) 04:52, 18 June 2026 (UTC)

duplicate article?

Hello, Please consider merging this article with https://en.wikipedia.org/wiki/Intelligent_agent ~2026-34907-40 (talk) 11:19, 15 June 2026 (UTC)

They are not the same topic. Jacob Rampino (talk) 04:51, 18 June 2026 (UTC)

Suggested addition to Further reading (COI disclosure)

Conflict of interest disclosure: I am the author of the book referenced below. Per Wikipedia:Conflict of interest, I am proposing this change here rather than editing the article directly.

Proposed addition: I'd like to suggest adding the following to the "Further reading" section:

• Nagasubramanian, Dhivya. Agentic AI for Engineers. Apress/Springer, 2026.

 ISBN 979-8-8688-2360-2. DOI: https://doi.org/10.1007/979-8-8688-2361-9

Why this is relevant: The book covers [topic — e.g., "agentic AI system design, safety, and enterprise deployment patterns"], which falls within the scope of this article.

Independent evidence of notability/relevance: ["The book has been adopted in 266 libraries across the country https://search.worldcat.org/title/1584648098 "

I'd appreciate any feedback from editors on whether this addition is appropriate. Happy to adjust the citation format or wording as needed.

Rine7986 (talk) 02:35, 20 June 2026 (UTC)

Thanks for disclosing. Did you notice that the article is tagged for having unreliable sources? Instead or promoting your own book, please help improve the article by proposing specific uses of this book to fix some of the article's fundamental problems.

I have set this request to answered. Please feel free to continue discussing, but do not re-open the request template until consensus has changed. Grayfell (talk) 02:38, 20 June 2026 (UTC)

Additions to the Overview paragraph

Wanted to suggest adding additional use cases -

Existing - "A common application of AI agents is task automation: for example, booking travel plans based on a user's prompted request.[7][8][9]. "

Suggestion - A common application of AI agents is task automation: for example, booking travel plans based on a user's prompted request.[7][8][9]. In the physical world, an AI agent could monitor real-time video and vision systems in a warehouse to identify events outside of normal operations. JPMorgan Chase is exploring the use of AI agents to detect fraud, provide customized financial advice. In areas that involve a lot of counterparties or that require a substantial effort to evaluate options — startup funding, college admissions, or B2B procurement, to name a few — agents deliver value by reading reviews, analyzing metrics, and comparing attributes across a range of options.

Existing: Companies such as Google, Microsoft and Amazon Web Services have offered platforms for deploying pre-built AI agents.[10] Several protocols have been proposed for standardizing inter-agent communication, such as the Model Context Protocol, Gibberlink,[11],

Suggestion - Companies such as Google, Microsoft and Amazon Web Services have offered platforms for deploying pre-built AI agents.[10] Several protocols have been proposed for standardizing inter-agent communication, such as the Model Context Protocol, Gibberlink,[11], NANDA Index , AGNTCY , A2A etc under the Linux foundation and AAIF umbrella. (all citations are here Sujata.tibre (talk) 00:43, 1 July 2026 (UTC))

Not sure what you were trying to do, but these proposals lack sources. The first is also written in an inappropriately conversational and vague tone. Please do not use LLMs to propose changes to Wikipedia articles. Thank you. Grayfell (talk) 00:22, 2 July 2026 (UTC)

The sources seem to have a formatting error, they are available on User talk:Sujata.tibre in a teahouse discussion. The second recommended addition seems reasonably well sourced and noteworthy for an encyclopedia. Jacob Rampino (talk) 00:41, 2 July 2026 (UTC)

A laundry list of overlapping protocols without any context is not useful to readers. But we can discuss more when actual sources are here, so they can be directly evaluated in context. Grayfell (talk) 01:20, 2 July 2026 (UTC)

The list of sources is here, the user seems new to Wikipedia and its markdown.

There are sources from ACM, Linux Foundation, MIT, Forbes, and WSJ, which would indicate that there are available reliable sources supporting some of the list, especially NANDA. Regarding WP:VNOT, one would probably need to look to other policies to explain why some protocols would be worth including and others not.

A protocol may need more contextualization (and obviously there are some grammatical issues with the list) but I wouldn't think that negates mention because one might characterize them as overlapping or in some way too exhaustive. Jacob Rampino (talk) 01:46, 2 July 2026 (UTC)

Unsurprisingly, this kind of thing comes up on Wikipedia very often.

Asking editors to check malformed refs on some other page is not sufficient. For consensus to form for inclusion here, the specific sources need to also be here for discussion.

At a glance, I am not at all convinced those sources are reliable in context. WP:PRIMARY, WP:FORBESCON, etc.

But my point remains. A list of names without context is not helpful. Both Model Context Protocol and Gibberlink have articles. (They are not necessarily good article, but they exist). NANDA Index, and AGNTCY do not (A2A is a false match). If the mere existence of these other protocols is informative, consider either WP:WTAF for that context or at least citing and summarizing more substantial, independent sources to provide useful information to readers, rather than providing inscrutable acronyms. Grayfell (talk) 03:37, 2 July 2026 (UTC)

Sorry I may have misunderstood, and this might help me with new editors and building consensus in the future.

Which policy that states For consensus to form for inclusion here, the specific sources need to also be here for discussion?

Or that Asking editors to check malformed refs on some other page is not sufficient?

I don’t see this mentioned in WP:V, for example. Certainly they weren’t malformed such that they couldn’t be discussed.

My point is that on some level the editor is clearly attempting to share refs, and is new. A link directly to where the editor gave them is providing sourcing well enough to evaluate them on the merits as was just done.

In either case, thank you for reviewing the sources. Jacob Rampino (talk) 04:13, 2 July 2026 (UTC)

At a glance, some of these sources are not reliable. Maybe some of them are reliable. Discussing all of them here, while linking to some other page which may eventually be edited or blanked, or archived so the link no longer works, will make this discussion even harder for others to follow than it already is. The goal of this talk page is to improve the article, but talk pages also act as records of the discussion. If you want to help this editor, I would suggest instead of defending these edits in the abstract based on vague principles, either take responsibility for the edits as actual edits in a way that other people can discuss, or communicate with that editor to allow them to make that case on their own. Make sure they know about the WP:TEAHOUSE, and H:PREVIEW, and so on. Grayfell (talk) 04:44, 2 July 2026 (UTC)

I do not understand what you mean by based on vague principles. I did not defend the edits aside from pointing out that the user was already seeking help on the teahouse and that engaging with the sources they provide is pertinent per WP:V when discussing inclusion.

Thank you for clarifying the policy as to sourcing on talk pages. Jacob Rampino (talk) 05:01, 2 July 2026 (UTC)

If your not defending the edits, than this has drifted very far off topic. Other editors are not expected to engage with the sources if they cannot find those sources. Further, I think you are confused about how Template:Help me and the WP:TEAHOUSE work, but again, this isn't the place to go in to that. This is instead the place to discussion actionable changes to the AI agent article. Grayfell (talk) 05:11, 2 July 2026 (UTC)

Hi @Grayfell & @Jacob Rampino I have been reading and re-reading this conversation, and I am not sure what are the bext steps for me. Here is what I am taking away 1) Correct the sources, (my bad one of them is not linking to the right place) 2) I took the use cases summary from the article I had linked there, because one of the wiki editiors I had met at UN edit-athon in New York told me every edit should have a reference 3) You are suggesting I create a page for AGNTCY, A2A and NANDA index since just adding them as is doesn't add value. Did i get the summary right ? Sujata.tibre (talk) 20:01, 8 July 2026 (UTC)

@Sujata.tibre: Hello, and welcome back. The editor you spoke with gave you good advice! Put simply, every edit which adds information should have a source.

Part of the problem is that your link to the sources you posted doesn't work. But it's clear you're trying to link to your own talk page. Since talk pages change at different rates, this is difficult or even impossible for future editors to understand. If your account is renamed, or your talk page is archived or deleted in the future, this discussion becomes incomprehensible.

You should usually post proposed sources to the article's talk page (meaning this talk page). Please also include a Template:reflist-talk at the bottom as well.

So, the sources you're proposing are here: Special:PermanentLink/1362106506... Right? Are those the sources you intended to post?

If so, I see some problems. Some of these are WP:PRIMARY sources, and should be avoided for minor details like this. Generally, articles should avoid press releases. We can use press releases, but only with a good reason, and that reason should either be really, really obvious, or should be supported by a reliable, independent source. Neither of those apply here. Also, at least one of them isn't reliable at all, since it's a WP:FORBESCON blog post.

The deeper issue is that the article should provide context for the benefit of disinterested readers. Mentioning the existence of these protocols isn't helpful to readers without context, and that context should come from reliable, independent sources.

If amultiple reliable, independent sources exist for NANDA Index, AGNTCY, or A2A protocol, they might be notable enough for their own articles, in which case this would be a reasonable place to add links.

But obviously, writing a new article from scratch is a much larger project.

I hope that answers your questions. Grayfell (talk) 22:02, 8 July 2026 (UTC)

Thanks @Grayfell, so as per definitions which one of the following are admissible? based on that I will find other sources and see if they can be made into a page each, or we can decide that cannot be inducted into wikipedia yet. I understand that no. 4 is not admissible given it is forbes con, but I am still confused about primary sources. I can see that no 2 below woule be primary since prof ramesh is from MIT media labs, but he is not from SLOAN, so would 1 be still considered primary? How about 3, 5, 6 and the 3rd party sources below are those admissible, If you can guide me here it will be extremely helpful.

1. MIT SLOAN : https://mitsloan.mit.edu/ideas-made-to-matter/agentic-ai-explained (NANDA), 2. 2. 2. MIT media Labs: https://www.media.mit.edu/projects/mit-nanda/overview/ (primary as Pr

3. ACM : https://dl.acm.org/doi/abs/10.1145/3774905.3793921?__cf_chl_f_tk=3dLgtbexvVq7PcS6H.Sj5B7Qoif4GQUZ_QaoaRMfFTo-1782865415-1.0.1.1-6ZL6SxUVHBcg_dofCtqTTQUNEO3rYF2rCacOL4w6Rls (peer reviewed independant Journal that mentions AGNTCY, NANDA and A2A

4. Forbes : https://www.forbes.com/sites/johnwerner/2025/04/29/theyre-making-tcpip-for-ai-and-its-called-nanda/ (Is this FORBES con)

5. Newstack: https://thenewstack.io/how-mits-project-nanda-aims-to-decentralize-ai-agents/

6. Linux Foundation

1. AGNTCY: "Linux Foundation Welcomes the AGNTCY Project to Standardize Open Multi-Agent System Infrastructure and Break Down AI Agent Silos".
2. A2A: https://www.linuxfoundation.org/press/linux-foundation-launches-the-agent2agent-protocol-project-to-enable-secure-intelligent-communication-between-ai-agents
3. A2A: https://www.linuxfoundation.org/press/a2a-protocol-surpasses-150-organizations-lands-in-major-cloud-platforms-and-sees-enterprise-production-use-in-first-year

3rd Party:

AGNTCY-

7.Dell : https://www.dell.com/en-us/blog/dell-drives-open-source-ai-innovation-with-agntcy-linux-foundation/

8. Galileo: https://galileo.ai/blog/agntcy-open-collective-multi-agent-standardization

9.Venture Beat : https://venturebeat.com/ai/a-standard-open-framework-for-building-ai-agents-is-coming-from-cisco-langchain-and-galileo

10. Wall street journal: https://partners.wsj.com/cisco/building-the-internet-of-agents/why-we-need-a-new-internet-for-ai/

A2A

12. Cyber Risc Alliance : https://www.scworld.com/resource/from-m2m-to-a2a-why-your-identity-stack-wasnt-built-for-whats-coming

13. Oracle: https://blogs.oracle.com/fusioncoe/fusion-ai-agents-a2a-teaching-your-ai-agents-to-actually-talk-to-each-other

AGNTCY & A2A

14. Oracle: https://blogs.oracle.com/ai-and-datascience/agentic-ai-foundation Sujata.tibre (talk) 19:54, 20 July 2026 (UTC)

This is much easier to discuss, thanks.

So I have two separate concerns here.

The first concern I have is that most of these sources are not reliable in this context.

The Wall Street Journal is an "advertorial"" from Cisco. Likewise, blogs from corporations announcing that they are "committed" to using a new standard are not 3rd party sources, nor are they reliable sources. Blogs in general are not reliable (per WP:SPS), and these blogs specifically are not reliable, because they are functionally advertising. Similarly, press releases generally make for poor quality sources, so the Linux Foundations links are poor for this.

The first MIT Sloan link doesn't appear to mention NANDA, so it's not usable

That leaves 3, 5, 9, and 12. For reasons I don't think we need to get in to, I don't accept that 12 (the SCWorld "resource") is a reliable source. 9 (the Venture beat article) is from March 2025, when the protocol was first announced. It's very speculative, talking about the group's future intentions for the protocol, how they "still needs to recruit more" companies, and how "projects like AGNTCY face an uphill battle". Explaining AGNTCY with this source would need a lot of context. This brings us down to only two sources.

So now I will talk about the second concern I have:

I don't think naming these protocols is going to be useful to readers of this article. Name-dropping these projects doesn't tell readers anything of value. We need to be able to provide context, and the way to do that is via reliable, independent sources. The best way to do that is to write articles for these other protocols, but these sources you have proposed do not demonstrate notability.

So as proposed, I don't think this would be an improvement to the article.

Grayfell (talk) 21:02, 20 July 2026 (UTC)

Hello thanks so much, I really appreciate this coaching and this actually tells me what is relible and what is not, so I can take this learning to pick another topic to contribute and let this go. This space is anyway evolving and so perhaps this needs more time so they meet test of time, and I hope these can meet wiki standards given test of time... BTW does wiki use github community as reference? because open source moves very fast and community adoption (github stars, forks, contributions from independant developers) could be a good metrics to add. Not that these protocls would meet that right now, just asking for my own knowledge. Coming from the Industry I know of many protocols (not from agentic AI space, but other spaces) which are now defacto industry standards Sujata.tibre (talk) 17:54, 21 July 2026 (UTC)

Github repositories (or Hugging Face, etc.) are usually not reliable sources. They may meet WP:ABOUTSELF, but that would depend on context. The number of stars, forks, etc. is not significant without context from a reliable source. I do not accept that an exception should be made for this topic.

More broadly, keep in mind that Wikipedia is an encyclopedia, not a journal. We do not publish original research. Wikipedia may cover things on the cutting edge, but only if reliable sources support it. Otherwise such advances are outside of the project's scope. To put it another way, if a field is rapidly evolving, we still need to summarize based on reliable, independent sources. Grayfell (talk) 19:35, 21 July 2026 (UTC)

frontier agents - new class

Kiro, AWS Security Agen, AWS DevOps Agent Idk if it is or isn't a new class of AI agents. ~2026-38073-05 (talk) 06:48, 4 July 2026 (UTC)

It's yet another marketing jargon. Same old same old. It's like marketing a spade as a manual geomorphological modification implement, but it's still a spade. 海盐沙冰 _{/ aka irisChronomia / Talk} 06:57, 4 July 2026 (UTC)

maybe. the fact is Amazon Kiro is replacing Amazon Q Developer and if you read about it, it is quite different. Microsoft/GitHub has sth similar I am not so sure it's just a marketing jargon. From what I could gather it seems those are "bare bones" tools/agents that are deployed to companies (via frontier engineers) to develop their own internal agents. So the point is to allow companies to develop their own TeslaOne (there is a whole story about it and how it gives Tesla an edge, now Honda and Toyota are developing their own, but the important fact is it took them 10+ years to get there). So basically it seems those are barebones agents that companies develop/train internally and the benefit in this case for Amazon is that it's used on AWS and for the company the benefit is it doesn't have knowledge it's unnecesarry for them and it's trained on their data. So it's kind of local but not completely. I could be way oof mark but it doesn't seem it's just jargon speak.~2026-38073-05 (talk) 07:15, 4 July 2026 (UTC)

it may be very different, i don't doubt the possibility, (though i'm also the least of a fan of LLMs), and we may keep an eye out for mainstream coverages. currently it feels like something WikiNews would consider, since it's just a product announcement. 海盐沙冰 _{/ aka irisChronomia / Talk} 07:21, 4 July 2026 (UTC)

my problem with those corporations' way of doing announcements is that they don't offer much details about their implementation, in the sense that they don't even publish a paper describing the new ideas they're coming up with (e.g., unlike how AMD/Intel used to publish ICCSE papers everytime they do a new microarchitecture). So I really don't know what to write about them, other than "[corporation] announced [product], claimed [list of improvements]". 海盐沙冰 _{/ aka irisChronomia / Talk} 07:24, 4 July 2026 (UTC)

Stateless and Stateful/Persistent agents ?

It seems the distinction is whether AI agents use stateles or stateful runtime.
What Is a Stateful Runtime? , Stateful Agents Redefine Infrastructure Governance , Amazon and OpenAI Bet Big on Stateful AI Infrastructure , Amazon Bedrock Stateful Runtime: Build Persistent AI Agents , Introducing the Stateful Runtime Environment for Agents in Amazon Bedrock , Introducing stateful MCP client capabilities on Amazon Bedrock AgentCore Runtime (vs stateless) I think it's worth including. ~2026-38073-05 (talk) 07:03, 4 July 2026 (UTC)

Honestly, agents are just machine learning models (which happen to be mostly LLMs) hooked onto external APIs capable of performing some work. The rest are just details about specific LLM implementations. 海盐沙冰 _{/ aka irisChronomia / Talk} 07:07, 4 July 2026 (UTC)

yes it true. but there are differences. a lens is just a glass that focuses light. does that mean symmetrical and asymmetrical lenses shouldn't be covered in the lens article because on some macro level they are the same thing? Rose is just a plant, that is it's in the plantae kingdom - by that logic on macrolevel Rose article is unnecesarry. ~2026-38073-05 (talk) 07:22, 4 July 2026 (UTC)

I know, and I'm clearly biased against the LLM / machine learning companies here; I just think if these companies are really doing something different, their products would be covered by the wider media, instead of many press releases. 海盐沙冰 _{/ aka irisChronomia / Talk} 07:27, 4 July 2026 (UTC)

Agreed. None of those sources are usable. Grayfell (talk) 07:31, 4 July 2026 (UTC)