Curious to know what your experience has been and the pros and cons of each one."},"title":{"fullyHighlighted":true,"matchLevel":"full","matchedWords":["ask","hn","best","multi","agent","frameworks"],"value":"Ask HN: Best Multi-Agent Frameworks?"}},"_tags":["story","author_digitcatphd","story_39742800","ask_hn"],"author":"digitcatphd","children":[39875282],"created_at":"2024-03-18T11:48:38Z","created_at_i":1710762518,"num_comments":1,"objectID":"39742800","points":22,"story_id":39742800,"story_text":"Does anyone have experience testing the various multi-agent frameworks (E.g. AutoGen, CrewAI, Agency Swarm)
Curious to know what your experience has been and the pros and cons of each one.","title":"Ask HN: Best Multi-Agent Frameworks?","updated_at":"2024-09-20T16:37:37Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"bberenberg"},"story_text":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["multi","agent","frameworks"],"value":"1. Has anyone recently compared the various OSS multi agent frameworks and have strong opinions on what to start with and why? I have a preference for code based, but a UI to visualize would be nice.\n - Phitdata\n - AgentChat 0.4(RC)\n - Crew\n - Other?
2. In general, what is the most "batteries included" way to work with agents? My goal is to reduce time spent integrating tools (logging, monitoring, evaluation, security, etc) while I build PoCs. My current approach is LiteLLM + LangFuse, but would love to head about something better."},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ask","hn","best","multi","agent"],"value":"Ask HN: What's the best way to build multi agent PoCs today"}},"_tags":["story","author_bberenberg","story_42331784","ask_hn"],"author":"bberenberg","children":[42332226],"created_at":"2024-12-05T19:36:11Z","created_at_i":1733427371,"num_comments":1,"objectID":"42331784","points":4,"story_id":42331784,"story_text":"1. Has anyone recently compared the various OSS multi agent frameworks and have strong opinions on what to start with and why? I have a preference for code based, but a UI to visualize would be nice.\n - Phitdata\n - AgentChat 0.4(RC)\n - Crew\n - Other?
2. In general, what is the most "batteries included" way to work with agents? My goal is to reduce time spent integrating tools (logging, monitoring, evaluation, security, etc) while I build PoCs. My current approach is LiteLLM + LangFuse, but would love to head about something better.","title":"Ask HN: What's the best way to build multi agent PoCs today","updated_at":"2024-12-05T23:02:51Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jdw64"},"story_text":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["best","multi","agent","frameworks"],"value":"I am a programmer with about seven years of experience, but I would say I am below many of the programmers here, and most of the work I do is fairly simple. I mainly work on WPF and WinForms applications that act as UI layers for ladder diagram based systems in industrial sites.
Because of the nature of this work, it is very labor intensive. There are too many factory trips, it is physically exhausting, and the pay is not great relative to the amount of effort involved. That is why I want to become more of a product oriented programmer. To move in that direction, I have been trying to use AI as actively as possible. In my current job, though, that is difficult because factory security is usually closed off and AI tools cannot really be connected there.
I have been reading about prompt engineering, harness style documentation, and similar topics. I also use OpenClaw, and honestly I feel like I am already trying almost everything I can get my hands on. I connect it with Obsidian and write down the knowledge I think my agents need while I work.
Still, the idea of ten times productivity feels somewhat exaggerated to me. I want to know how people actually get better at using AI, and how they learn the underlying methods. There is so much hype around AI that it is hard to tell what is real, and learning this stuff has been more difficult than I expected.
How should I study this properly?
Right now I especially want to learn how to manage multi agent systems. Every AI only multi agent framework I have built or tried so far has failed. At first I tried to control it in a TDD like way, but people who have used TDD seriously probably know what I mean when I say that tests can become too locally focused. Sometimes the architecture starts to fall apart, and then the agents keep fixing only those small areas over and over while the token cost keeps rising.
At the same time, in Korea there is a huge amount of talk that if you are not using AI, you will fall behind. Because of that, I have been trying hard to learn it so I do not get left behind. And to be honest, programming has become much more enjoyable for me since I started using AI.
One reason is that programming feels deeply tied to English ways of thinking, and that has always felt awkward with Korean. Even the act of writing code used to feel mentally heavy for me. But with AI, I can think through things in Korean and still code effectively.
You know that huge fatigue you feel when you first start writing code from scratch?
When I write a single interface, I immediately start seeing the number of implementations.\nWhen I see the implementations, I start seeing lifecycle conflicts.\nWhen I see lifecycle issues, I start thinking about ownership and disposal.\nWhen I think about ownership, I start thinking about pooling possibilities and reset contracts.\nWhen I think about DI, I start seeing the composition root and the test seams.\nWhen I think about the contract itself, I start worrying about future extension costs.
All of that used to make coding feel painfully heavy for me.
But AI just writes a draft without getting stuck in all of that, and I genuinely enjoy taking that draft and reshaping it around my own thinking.
I want to get much better at using AI this way.
What are the best ways to improve at it, and how do you keep up with useful trends without getting buried in the hype?"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ask","hn"],"value":"Ask HN: May be a basic question, but how can I use AI well?"}},"_tags":["story","author_jdw64","story_47822787","ask_hn"],"author":"jdw64","children":[47826541,47829540,47831198,47842010,47843519,47862377,47870658],"created_at":"2026-04-19T08:42:37Z","created_at_i":1776588157,"num_comments":5,"objectID":"47822787","points":10,"story_id":47822787,"story_text":"I am a programmer with about seven years of experience, but I would say I am below many of the programmers here, and most of the work I do is fairly simple. I mainly work on WPF and WinForms applications that act as UI layers for ladder diagram based systems in industrial sites.
Because of the nature of this work, it is very labor intensive. There are too many factory trips, it is physically exhausting, and the pay is not great relative to the amount of effort involved. That is why I want to become more of a product oriented programmer. To move in that direction, I have been trying to use AI as actively as possible. In my current job, though, that is difficult because factory security is usually closed off and AI tools cannot really be connected there.
I have been reading about prompt engineering, harness style documentation, and similar topics. I also use OpenClaw, and honestly I feel like I am already trying almost everything I can get my hands on. I connect it with Obsidian and write down the knowledge I think my agents need while I work.
Still, the idea of ten times productivity feels somewhat exaggerated to me. I want to know how people actually get better at using AI, and how they learn the underlying methods. There is so much hype around AI that it is hard to tell what is real, and learning this stuff has been more difficult than I expected.
How should I study this properly?
Right now I especially want to learn how to manage multi agent systems. Every AI only multi agent framework I have built or tried so far has failed. At first I tried to control it in a TDD like way, but people who have used TDD seriously probably know what I mean when I say that tests can become too locally focused. Sometimes the architecture starts to fall apart, and then the agents keep fixing only those small areas over and over while the token cost keeps rising.
At the same time, in Korea there is a huge amount of talk that if you are not using AI, you will fall behind. Because of that, I have been trying hard to learn it so I do not get left behind. And to be honest, programming has become much more enjoyable for me since I started using AI.
One reason is that programming feels deeply tied to English ways of thinking, and that has always felt awkward with Korean. Even the act of writing code used to feel mentally heavy for me. But with AI, I can think through things in Korean and still code effectively.
You know that huge fatigue you feel when you first start writing code from scratch?
When I write a single interface, I immediately start seeing the number of implementations.\nWhen I see the implementations, I start seeing lifecycle conflicts.\nWhen I see lifecycle issues, I start thinking about ownership and disposal.\nWhen I think about ownership, I start thinking about pooling possibilities and reset contracts.\nWhen I think about DI, I start seeing the composition root and the test seams.\nWhen I think about the contract itself, I start worrying about future extension costs.
All of that used to make coding feel painfully heavy for me.
But AI just writes a draft without getting stuck in all of that, and I genuinely enjoy taking that draft and reshaping it around my own thinking.
I want to get much better at using AI this way.
What are the best ways to improve at it, and how do you keep up with useful trends without getting buried in the hype?","title":"Ask HN: May be a basic question, but how can I use AI well?","updated_at":"2026-06-03T14:37:14Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"szemy2"},"story_text":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["hn","multi","agent","frameworks"],"value":"Hi HN!
I am currently interested in learning about Complex Systems and the theory that has been developed in the recent decades.
I am reading the excellent book called 'Introduction to the Theory of Complex Systems' (https://www.amazon.de/Introduction-Theory-Complex-Systems-Thurner/dp/019882193X) which rigorously goes through the fundamentals on the theoretical framework.
It is however not very hands-on: I would like to code simulations on the side and get my hands dirty with coding.
I am looking for a course that specifically targets algorithms/frameworks that is used by Complexity Researchers. I know complexity sciences touches many disciplines, but the premise of the book is that there are underlying patterns in all types of complex systems, so a generalist approach would be preferable. (I am not interested in beginner course in coding, for eg.: Nature of Code (https://www.youtube.com/user/shiffman) touches on some algorithms to simulate multi-agent behaviour)
Please point me towards a good source :)"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ask","hn","best"],"value":"Ask HN: Best Programming Course for Complex Systems?"}},"_tags":["story","author_szemy2","story_22281471","ask_hn"],"author":"szemy2","children":[22283616],"created_at":"2020-02-09T10:47:26Z","created_at_i":1581245246,"num_comments":3,"objectID":"22281471","points":2,"story_id":22281471,"story_text":"Hi HN!
I am currently interested in learning about Complex Systems and the theory that has been developed in the recent decades.
I am reading the excellent book called 'Introduction to the Theory of Complex Systems' (https://www.amazon.de/Introduction-Theory-Complex-Systems-Thurner/dp/019882193X) which rigorously goes through the fundamentals on the theoretical framework.
It is however not very hands-on: I would like to code simulations on the side and get my hands dirty with coding.
I am looking for a course that specifically targets algorithms/frameworks that is used by Complexity Researchers. I know complexity sciences touches many disciplines, but the premise of the book is that there are underlying patterns in all types of complex systems, so a generalist approach would be preferable. (I am not interested in beginner course in coding, for eg.: Nature of Code (https://www.youtube.com/user/shiffman) touches on some algorithms to simulate multi-agent behaviour)
Please point me towards a good source :)","title":"Ask HN: Best Programming Course for Complex Systems?","updated_at":"2024-09-20T05:43:44Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"kken"},"story_text":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["best"],"value":"The projects mentioned in the title have garnedered an impressive amount of "stars" in Github. The concept is really compelling, but whenever I tried them on a real problem I ended up with a heap of very verbose information and a codebase that usually did not work out of the box. It usually felt easier to work on the problem step by step.
Are there some inspiring descriptions of success stories? How to best use these tools?"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ask","hn","multi","agent","frameworks"],"value":"Ask HN: Success stories for Multi-Agent GPT frameworks like MetaGPT, GPTEngineer"}},"_tags":["story","author_kken","story_37109586","ask_hn"],"author":"kken","created_at":"2023-08-13T12:59:37Z","created_at_i":1691931577,"num_comments":0,"objectID":"37109586","points":2,"story_id":37109586,"story_text":"The projects mentioned in the title have garnedered an impressive amount of "stars" in Github. The concept is really compelling, but whenever I tried them on a real problem I ended up with a heap of very verbose information and a codebase that usually did not work out of the box. It usually felt easier to work on the problem step by step.
Are there some inspiring descriptions of success stories? How to best use these tools?","title":"Ask HN: Success stories for Multi-Agent GPT frameworks like MetaGPT, GPTEngineer","updated_at":"2024-09-20T14:45:58Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jiayuanzhang"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["ask","hn","best","multi","agent","frameworks"],"value":"Hi HN,
I am Jiayuan, and I'm here to introduce a tool we've been building over the past few months: Devv (https://devv.ai). In simple terms, it is an AI-powered search engine specifically designed for developers.
Now, you might ask, with so many AI search engines already available\u2014Perplexity, You.com, Phind, and several open-source projects\u2014why do we need another one?
We all know that Generative Search Engines are built on RAG (Retrieval-Augmented Generation)[1] combined with Large Language Models (LLMs). Most of the products mentioned above use indexes from general search engines (like Google/Bing APIs), but we've taken a different approach.
We've created a vertical search index focused on the development domain, which includes:
- Documents: These are essentially the single source of truth for programming languages or libraries; I believe many of you are users of Dash (https://kapeli.com/dash) or devdocs (https://devdocs.io/).
- Code: While not natural language, code contains rich contextual information. If you have a question related to the Django framework, nothing is more convincing than code snippets from Django's repository.
- Web Search: We still use data from search engines because these results contain additional contextual information.
Our reasons for doing this include:
- The quality of the index is crucial to the RAG system; its effectiveness determines the output quality of the entire system.
- We focus more on the Index (RAG) rather than LLMs because LLMs evolve rapidly; even models performing well today may be superseded by better ones in a few months, and fine-tuning an LLM now has relatively low costs.
- All players are currently exploring what kind of LLM product works best; we hope to contribute some different insights ourselves (and plan to open source parts of our underlying infrastructure in return for contributions back into open source communities).
Some brief product features:
- Three modes: - Fast mode: Offers quick answers within seconds. - Agent mode: For complex queries where Devv Agent infers your question before selecting appropriate solutions. - GitHub mode(currently in beta): Links directly with your own GitHub repositories allowing inquiries about specific codebases.
- Clean & intuitive UI/UX design.
- Currently only available as web version but Chrome extension & VSCode plugin planned soon!
Technical details regarding how we build our Index:
- Documents section involves crawling most documentation sources using scripts inspired by devdocs project\u2019s crawler logic then slicing them up according function/symbol dimensions before embedding into vector databases;
- Codes require special treatment beyond just embeddings alone hence why custom parsers were developed per language type extracting logical structures within repos such as architectural layouts calling relationships between functions definitions etc., semantically processed via LMM;
- Web searches combine both selfmade indices targeting developer niches alongside traditional API based methods. We crawled relevant sites including blogs forums tech news outlets etc..
For the Agent Mode, we have actually developed a multi-agent framework. It first categorizes the user's query and then selects different agents based on these categories to address the issues. These various agents employ different models and solution steps.
Future Plans:
- Build a more comprehensive index that includes internal context (The Devv for Teams version will support indexing team repositories, documents, issue trackers for Q&A)
- Fully localized: All of the above technologies can be executed locally, ensuring privacy and security through complete localization.
Devv is still in its very early stages and can be used without logging in. We welcome everyone to experience it and provide feedback on any issues; we will continue to iterate on it.
[1]: https://arxiv.org/abs/2005.11401"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["hn"],"value":"Show HN: I made a better Perplexity for developers"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://devv.ai"}},"_tags":["story","author_jiayuanzhang","story_40299091","show_hn"],"author":"jiayuanzhang","children":[40299769,40299784,40299801,40299817,40299859,40299862,40299918,40299988,40300118,40300280,40300313,40300341,40300360,40300366,40300558,40300629,40300771,40300874,40300924,40301120,40301299,40301862,40301950,40302461,40304085,40304263,40309281,40315256],"created_at":"2024-05-08T15:19:40Z","created_at_i":1715181580,"num_comments":75,"objectID":"40299091","points":185,"story_id":40299091,"story_text":"Hi HN,
I am Jiayuan, and I'm here to introduce a tool we've been building over the past few months: Devv (https://devv.ai). In simple terms, it is an AI-powered search engine specifically designed for developers.
Now, you might ask, with so many AI search engines already available\u2014Perplexity, You.com, Phind, and several open-source projects\u2014why do we need another one?
We all know that Generative Search Engines are built on RAG (Retrieval-Augmented Generation)[1] combined with Large Language Models (LLMs). Most of the products mentioned above use indexes from general search engines (like Google/Bing APIs), but we've taken a different approach.
We've created a vertical search index focused on the development domain, which includes:
- Documents: These are essentially the single source of truth for programming languages or libraries; I believe many of you are users of Dash (https://kapeli.com/dash) or devdocs (https://devdocs.io/).
- Code: While not natural language, code contains rich contextual information. If you have a question related to the Django framework, nothing is more convincing than code snippets from Django's repository.
- Web Search: We still use data from search engines because these results contain additional contextual information.
Our reasons for doing this include:
- The quality of the index is crucial to the RAG system; its effectiveness determines the output quality of the entire system.
- We focus more on the Index (RAG) rather than LLMs because LLMs evolve rapidly; even models performing well today may be superseded by better ones in a few months, and fine-tuning an LLM now has relatively low costs.
- All players are currently exploring what kind of LLM product works best; we hope to contribute some different insights ourselves (and plan to open source parts of our underlying infrastructure in return for contributions back into open source communities).
Some brief product features:
- Three modes: - Fast mode: Offers quick answers within seconds. - Agent mode: For complex queries where Devv Agent infers your question before selecting appropriate solutions. - GitHub mode(currently in beta): Links directly with your own GitHub repositories allowing inquiries about specific codebases.
- Clean & intuitive UI/UX design.
- Currently only available as web version but Chrome extension & VSCode plugin planned soon!
Technical details regarding how we build our Index:
- Documents section involves crawling most documentation sources using scripts inspired by devdocs project\u2019s crawler logic then slicing them up according function/symbol dimensions before embedding into vector databases;
- Codes require special treatment beyond just embeddings alone hence why custom parsers were developed per language type extracting logical structures within repos such as architectural layouts calling relationships between functions definitions etc., semantically processed via LMM;
- Web searches combine both selfmade indices targeting developer niches alongside traditional API based methods. We crawled relevant sites including blogs forums tech news outlets etc..
For the Agent Mode, we have actually developed a multi-agent framework. It first categorizes the user's query and then selects different agents based on these categories to address the issues. These various agents employ different models and solution steps.
Future Plans:
- Build a more comprehensive index that includes internal context (The Devv for Teams version will support indexing team repositories, documents, issue trackers for Q&A)
- Fully localized: All of the above technologies can be executed locally, ensuring privacy and security through complete localization.
Devv is still in its very early stages and can be used without logging in. We welcome everyone to experience it and provide feedback on any issues; we will continue to iterate on it.
[1]: https://arxiv.org/abs/2005.11401","title":"Show HN: I made a better Perplexity for developers","updated_at":"2025-03-17T23:27:01Z","url":"https://devv.ai"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jiayuanzhang"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["ask","hn","best","multi","agent","frameworks"],"value":"Hi HN,
I am Jiayuan, and I'm here to introduce a tool we've been building over the past few months: Devv (https://devv.ai). In simple terms, it is an AI-powered search engine specifically designed for developers.
Now, you might ask, with so many AI search engines already available\u2014Perplexity, You.com, Phind, and several open-source projects\u2014why do we need another one?
We all know that Generative Search Engines are built on RAG (Retrieval-Augmented Generation)[1] combined with Large Language Models (LLMs). Most of the products mentioned above use indexes from general search engines (like Google/Bing APIs), but we've taken a different approach.
We've created a vertical search index focused on the development domain, which includes:
- Documents: These are essentially the single source of truth for programming languages or libraries; I believe many of you are users of Dash (https://kapeli.com/dash) or devdocs (https://devdocs.io/).
- Code: While not natural language, code contains rich contextual information. If you have a question related to the Django framework, nothing is more convincing than code snippets from Django's repository.
- Web Search: We still use data from search engines because these results contain additional contextual information.
Our reasons for doing this include:
- The quality of the index is crucial to the RAG system; its effectiveness determines the output quality of the entire system.
- We focus more on the Index (RAG) rather than LLMs because LLMs evolve rapidly; even models performing well today may be superseded by better ones in a few months, and fine-tuning an LLM now has relatively low costs.
- All players are currently exploring what kind of LLM product works best; we hope to contribute some different insights ourselves (and plan to open source parts of our underlying infrastructure in return for contributions back into open source communities).
Some brief product features:
- Three modes:\n - Fast mode: Offers quick answers within seconds.\n - Agent mode: For complex queries where Devv Agent infers your question before selecting appropriate solutions.\n - GitHub mode(currently in beta): Links directly with your own GitHub repositories allowing inquiries about specific codebases.
- Clean & intuitive UI/UX design.
- Currently only available as web version but Chrome extension & VSCode plugin planned soon!
Technical details regarding how we build our Index:
- Documents section involves crawling most documentation sources using scripts inspired by devdocs project\u2019s crawler logic then slicing them up according function/symbol dimensions before embedding into vector databases;
- Codes require special treatment beyond just embeddings alone hence why custom parsers were developed per language type extracting logical structures within repos such as architectural layouts calling relationships between functions definitions etc., semantically processed via LMM;
- Web searches combine both selfmade indices targeting developer niches alongside traditional API based methods. We crawled relevant sites including blogs forums tech news outlets etc..
For the Agent Mode, we have actually developed a multi-agent framework. It first categorizes the user's query and then selects different agents based on these categories to address the issues. These various agents employ different models and solution steps.
Future Plans:
- Build a more comprehensive index that includes internal context (The Devv for Teams version will support indexing team repositories, documents, issue trackers for Q&A)
- Fully localized: All of the above technologies can be executed locally, ensuring privacy and security through complete localization.
Devv is still in its very early stages and can be used without logging in. We welcome everyone to experience it and provide feedback on any issues; we will continue to iterate on it.
[1]: https://arxiv.org/abs/2005.11401"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["hn"],"value":"Show HN: Devv \u2013 AI search engine for devs, built on a custom search index"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://devv.ai"}},"_tags":["story","author_jiayuanzhang","story_40288111","show_hn"],"author":"jiayuanzhang","children":[40288297,40288660,40288890,40290495,40315880],"created_at":"2024-05-07T16:46:18Z","created_at_i":1715100378,"num_comments":7,"objectID":"40288111","points":6,"story_id":40288111,"story_text":"Hi HN,
I am Jiayuan, and I'm here to introduce a tool we've been building over the past few months: Devv (https://devv.ai). In simple terms, it is an AI-powered search engine specifically designed for developers.
Now, you might ask, with so many AI search engines already available\u2014Perplexity, You.com, Phind, and several open-source projects\u2014why do we need another one?
We all know that Generative Search Engines are built on RAG (Retrieval-Augmented Generation)[1] combined with Large Language Models (LLMs). Most of the products mentioned above use indexes from general search engines (like Google/Bing APIs), but we've taken a different approach.
We've created a vertical search index focused on the development domain, which includes:
- Documents: These are essentially the single source of truth for programming languages or libraries; I believe many of you are users of Dash (https://kapeli.com/dash) or devdocs (https://devdocs.io/).
- Code: While not natural language, code contains rich contextual information. If you have a question related to the Django framework, nothing is more convincing than code snippets from Django's repository.
- Web Search: We still use data from search engines because these results contain additional contextual information.
Our reasons for doing this include:
- The quality of the index is crucial to the RAG system; its effectiveness determines the output quality of the entire system.
- We focus more on the Index (RAG) rather than LLMs because LLMs evolve rapidly; even models performing well today may be superseded by better ones in a few months, and fine-tuning an LLM now has relatively low costs.
- All players are currently exploring what kind of LLM product works best; we hope to contribute some different insights ourselves (and plan to open source parts of our underlying infrastructure in return for contributions back into open source communities).
Some brief product features:
- Three modes:\n - Fast mode: Offers quick answers within seconds.\n - Agent mode: For complex queries where Devv Agent infers your question before selecting appropriate solutions.\n - GitHub mode(currently in beta): Links directly with your own GitHub repositories allowing inquiries about specific codebases.
- Clean & intuitive UI/UX design.
- Currently only available as web version but Chrome extension & VSCode plugin planned soon!
Technical details regarding how we build our Index:
- Documents section involves crawling most documentation sources using scripts inspired by devdocs project\u2019s crawler logic then slicing them up according function/symbol dimensions before embedding into vector databases;
- Codes require special treatment beyond just embeddings alone hence why custom parsers were developed per language type extracting logical structures within repos such as architectural layouts calling relationships between functions definitions etc., semantically processed via LMM;
- Web searches combine both selfmade indices targeting developer niches alongside traditional API based methods. We crawled relevant sites including blogs forums tech news outlets etc..
For the Agent Mode, we have actually developed a multi-agent framework. It first categorizes the user's query and then selects different agents based on these categories to address the issues. These various agents employ different models and solution steps.
Future Plans:
- Build a more comprehensive index that includes internal context (The Devv for Teams version will support indexing team repositories, documents, issue trackers for Q&A)
- Fully localized: All of the above technologies can be executed locally, ensuring privacy and security through complete localization.
Devv is still in its very early stages and can be used without logging in. We welcome everyone to experience it and provide feedback on any issues; we will continue to iterate on it.
[1]: https://arxiv.org/abs/2005.11401","title":"Show HN: Devv \u2013 AI search engine for devs, built on a custom search index","updated_at":"2024-09-20T17:04:42Z","url":"https://devv.ai"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"zchmael"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["ask","hn","best","multi","agent","frameworks"],"value":"Hey HN,
I\u2019m Zack, CEO at Averi AI, and we just released Synapse, a modular AI architecture we built to solve a problem we kept running into within the marketing ecosystem:
\u201cHow do you get domain-specific intelligence without trying to recreate GPT-4 from scratch?\u201d
The Problem
Most domain-specific AI tools (marketing, legal, ops, etc.) tend to fall into one of three camps:\nUse GPT-4/Claude as-is and rely on prompt engineering
Train a small model from scratch but lose general reasoning
Go full frontier model\u2026 and burn millions trying
We\u2019ve considered all three. None hit the mark.
Our Approach: Multi-Model + Human Routing
Synapse is our attempt at something better:\nA routing architecture that matches tasks with the best resource whether that\u2019s an LLM, a smaller domain model, or a vetted human expert
A way to balance specialization and scale, instead of choosing one
It powers our own domain-specific foundation model (AGM-2), and integrates GPT-4, Claude, and others alongside it. Tasks get routed based on complexity and type.
For example:\nA quick product description \u2192 routed to AGM-2
A cross-channel campaign brief \u2192 goes through Strategic Cortex + GPT-4
A nuanced brand tone rewrite \u2192 routed to a human expert
Under the Hood
Architecture:\nSynapse is structured around 5 specialized cognitive modules (we call them cortices):\nBrief Cortex: Disambiguates messy requests
Strategic Cortex: Maps business goals to tactical plans
Creative Cortex: Writes content tuned to brand voice
Performance Cortex: Weighs historical campaign data
Human Cortex: Escalates to our expert network when needed
Routing Logic:
Dual-track complexity scoring: LLM + heuristic analysis
Tasks run in one of 3 \u201cmodes\u201d: Express (quick), Standard, or Deep (multi-stage, may call a human)
Results fed back to improve future routing decisions
Training Data:
AGM-2 was trained on over ~2M marketing artifacts (positioning docs, campaigns, A/B test data, etc.)\nWe licensed real performance data and layered in structured messaging frameworks. It\u2019s not the biggest model, but it\u2019s trained with domain-native intent.
What Makes This Different
Rather than trying to force one model to do everything, Synapse behaves more like a strategist. It knows when to go fast, when to go deep, and when to ask for help.
We\u2019ve been running it in production for 3+ months.
It\u2019s shown strong gains in:
Brand tone consistency vs. GPT-4-only setups
Time-to-launch on full campaigns
Quality of briefs when humans are looped in
Try It + Read More
Demo (mention you're from HN and we'll get you right in): https://www.averi.ai/demo-sign-up
Technical overview: https://www.averi.ai/blog/averi-launches-synapse-a-new-ai-sy...
Open Questions We\u2019re Exploring
Specialist vs. generalist tradeoffs \u2014 When does our domain-trained AGM-2 outperform GPT-4? When doesn\u2019t it?
Human-in-the-loop scaling \u2014 How do you decide when to escalate to a human? We use ML for this but would love to hear other approaches.
Training data \u2014 What\u2019s the right mix of public vs. proprietary when building domain-specific datasets?
Would love feedback from anyone building domain AI systems, orchestration layers, or multi-agent workflows. AMA on routing logic, model behavior, or anything else.
Thanks!"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["hn"],"value":"Show HN: Synapse \u2013 Multi-model AI combining LLMs and humans for marketing output"}},"_tags":["story","author_zchmael","story_44747099","show_hn"],"author":"zchmael","created_at":"2025-07-31T16:12:49Z","created_at_i":1753978369,"num_comments":0,"objectID":"44747099","points":1,"story_id":44747099,"story_text":"Hey HN,
I\u2019m Zack, CEO at Averi AI, and we just released Synapse, a modular AI architecture we built to solve a problem we kept running into within the marketing ecosystem:
\u201cHow do you get domain-specific intelligence without trying to recreate GPT-4 from scratch?\u201d
The Problem
Most domain-specific AI tools (marketing, legal, ops, etc.) tend to fall into one of three camps:\nUse GPT-4/Claude as-is and rely on prompt engineering
Train a small model from scratch but lose general reasoning
Go full frontier model\u2026 and burn millions trying
We\u2019ve considered all three. None hit the mark.
Our Approach: Multi-Model + Human Routing
Synapse is our attempt at something better:\nA routing architecture that matches tasks with the best resource whether that\u2019s an LLM, a smaller domain model, or a vetted human expert
A way to balance specialization and scale, instead of choosing one
It powers our own domain-specific foundation model (AGM-2), and integrates GPT-4, Claude, and others alongside it. Tasks get routed based on complexity and type.
For example:\nA quick product description \u2192 routed to AGM-2
A cross-channel campaign brief \u2192 goes through Strategic Cortex + GPT-4
A nuanced brand tone rewrite \u2192 routed to a human expert
Under the Hood
Architecture:\nSynapse is structured around 5 specialized cognitive modules (we call them cortices):\nBrief Cortex: Disambiguates messy requests
Strategic Cortex: Maps business goals to tactical plans
Creative Cortex: Writes content tuned to brand voice
Performance Cortex: Weighs historical campaign data
Human Cortex: Escalates to our expert network when needed
Routing Logic:
Dual-track complexity scoring: LLM + heuristic analysis
Tasks run in one of 3 \u201cmodes\u201d: Express (quick), Standard, or Deep (multi-stage, may call a human)
Results fed back to improve future routing decisions
Training Data:
AGM-2 was trained on over ~2M marketing artifacts (positioning docs, campaigns, A/B test data, etc.)\nWe licensed real performance data and layered in structured messaging frameworks. It\u2019s not the biggest model, but it\u2019s trained with domain-native intent.
What Makes This Different
Rather than trying to force one model to do everything, Synapse behaves more like a strategist. It knows when to go fast, when to go deep, and when to ask for help.
We\u2019ve been running it in production for 3+ months.
It\u2019s shown strong gains in:
Brand tone consistency vs. GPT-4-only setups
Time-to-launch on full campaigns
Quality of briefs when humans are looped in
Try It + Read More
Demo (mention you're from HN and we'll get you right in): https://www.averi.ai/demo-sign-up
Technical overview: https://www.averi.ai/blog/averi-launches-synapse-a-new-ai-sy...
Open Questions We\u2019re Exploring
Specialist vs. generalist tradeoffs \u2014 When does our domain-trained AGM-2 outperform GPT-4? When doesn\u2019t it?
Human-in-the-loop scaling \u2014 How do you decide when to escalate to a human? We use ML for this but would love to hear other approaches.
Training data \u2014 What\u2019s the right mix of public vs. proprietary when building domain-specific datasets?
Would love feedback from anyone building domain AI systems, orchestration layers, or multi-agent workflows. AMA on routing logic, model behavior, or anything else.
Thanks!","title":"Show HN: Synapse \u2013 Multi-model AI combining LLMs and humans for marketing output","updated_at":"2025-07-31T16:17:44Z"}],"hitsPerPage":20,"nbHits":8,"nbPages":1,"page":0,"params":"query=Ask+HN+Best+Multi-Agent+Frameworks&advancedSyntax=true&analyticsTags=backend","processingTimeMS":32,"processingTimingsMS":{"_request":{"roundTrip":16},"afterFetch":{"format":{"highlighting":1,"total":1}},"fetch":{"query":9,"scanning":21,"total":31},"total":32},"query":"Ask HN Best Multi-Agent Frameworks","serverTimeMS":34}