this is absolute trash how i used to read repos clone. open. scroll. guess. repeat an hour later - maybe i understand the structure then i found gitdiagram replace hub with diagram in any github url github.com/owner/repo → gitdiagram.com/owner/repo two minutes interactive architecture map click node → source code export png or mermaid ai video walkthrough included what it maps: frontend → api → services → storage → llm providers ai generated top level view not perfect static analysis but enough to know where to start stop reading folders start navigating

best practical guide on jev i have seen covers the full setup from playground to production two things stuck the shift: an llm creates the work jev decides what happens next that split changes the entire agent stack the menu rule: browser use rebuilds the action list after every click your chief of staff must rebuild the worker list after every tool call stale menu = stale decisions

came across a video on combining jev + claude code it builds a cheap agentic coding loop that actually holds up in real development → give the agent a clear task → it explores the codebase → makes the required changes → runs tests → returns a diff and a report → you review it, launch the next iteration don't ask ai to build the entire project break the work into small tasks each with its own acceptance criteria that one rule is what makes the loop survive contact with a real codebase works the same way for research infra, data pipelines, backtesting, trading tools - anywhere the task is too big to verify in one shot

every task on your stack runs through the same model typo fixes and deep synthesis get the same compute bill that is the waste nobody notices until the bill does i am implementing a model router using jev as the dispatcher three buckets: - fast track: typos, json formatting, basic extraction → llama 3 8b or phi-3 - standard track: short posts, summarization, simple contradictions → mistral or mid-size models - reasoning track: deep synthesis, strategic planning, complex code → gemma 4 31b or claude 3.5 jev reads the query picks the bucket routes it latency drops load drops the real test is not the router it is whether output quality holds once the cheap models take the easy half

too much ai noise every link feels important every thread feels like "alpha" my research agent was getting overwhelmed too many tokens spent on things that did not matter i needed a gatekeeper so i built a filter using JEV a system one model that does not think or summarize it just decides: yes or no the pizza test: > pepperoni pizza recipe: rejected > technical doc on agent architecture: approved do not let your most expensive model do the job of a simple filter use a fast, cheap decision layer to clean the input then let the researcher do the heavy lifting fewer tokens burned on things that never mattered

most use one big llm for everything it researches, it routes, it writes, it checks but a model good at writing is rarely good at deciding i am adding jev as a dedicated decision layer built into my hermes + ollama stack it never chats it never generates it only decides the pipeline: - input filter: scan links and drop the noise - quality gate: verify research briefs before writing - format router: decide between short post or deep article less waste faster loops that is a harness upgrade not a model upgrade

yesterday i told you about the hermes plugin catalog today i built my own stack 10 plugins that actually matter for my work search and research > web-search-plus: deep web research mode > hermes-rss: internal rss/atom reader > hermes-newswire: built-in news feed memory and optimization > yantrikdb: persistent cross-session memory > sugar: local memory for repetitive coding > tool-slimmer: cuts context waste on tool descriptions control and quality > grill-tab: task clarification and brief confirmation > diff-review: auto-scan git diffs for secrets and bugs > hermes-plugin-guard: security check before install > hermes-telemetry: local tracking of tokens and latency stop hunting for plugins start building a stack hermes-agent.nousresearch.co…

hermes ships 4 official plugins the community has added 96 that ratio is the interesting part the core team is not the bottleneck anymore someone needed desktop mods someone else needed a new platform someone built advanced browsing every community plugin gets reviewed new ones land regularly the shift is quiet: you stop waiting for the core update you install the tool you need an agent stops being a product once other people can extend it

most think ai filters are walls you set a rule, the model obeys then someone runs a role-play attack and the whole safety layer folds a filter is a request, not a barrier if your workflow depends on a model choosing to refuse the wrong thing you do not have a control you have a hope controls live outside the model what the agent can touch what it can run who signs off before send running local changes a different thing you own the stack no surprise update rewrites your workflow overnight that is a stability argument not a safety one stop tuning prompts to behave start building systems that do not need the model to

stop writing massive system prompts they fail in three ways: > they drift in long chats > they are a pain to copy-paste > they confuse the model with too many rules i solved this with "skills" procedural memory for agents a prompt is a whisper in the ear a skill is a manual on the desk how it works: i call a skill hermes loads the specific SKILL.md the model executes the exact algorithm the instructions are then cleared modularity: 100 different skills one for code review one for research one for writing no confusion. no drift if the agent fails, i do not re-prompt i patch the SKILL.md that is the difference: tuning a model versus building a library of thinking processes

this vortex is openai's proposed answer to the navier-stokes millennium problem $1 million still listed as open by clay not from one prompt roughly 10k concurrent agents 88 hours to a result 130 billion tokens in that run then 17 more hours to formalize and check it in lean nobody outside openai has finished the review openai says it will not even claim the prize the useful question is not "did the swarm win" it is whether the proof survives outside review

i installed a local model because i wanted to stop paying for every small AI task but that is not the main reason i keep using it a local model makes experimenting cheap you can give it bad prompts break a workflow rewrite a skill run the same task ten times without every message becoming another API bill it does not replace the best cloud models it just makes the learning loop much easier to afford

most people use AI as a chat box they ask a question, get an answer, and repeat i wanted a teammate one that knows how i think, not just what i ask so i built hermes no complex infrastructure no massive codebase just: hermes > ollama > custom skills > a stop-slop filter i did not aim for a smarter agent i aimed for a more honest one fed it my golden standard posts my failed experiments the styles of people i actually respect now it does not hallucinate a professional tone it knows when to be a minimalist when to be a curator when to be a story-hacker it does not write for me it filters the ai-slop out of my thoughts the real work was not the install it was the calibration an agent is only as useful as the constraints you give it testing next whether it can turn a raw technical log into a high-signal post without me touching a single word

today i did not build another AI agent i just kept working on the first one hermes + ollama + a local model + one researcher skill it is still small but it can already take a text and separate: > what was actually said > what needs evidence > what is still unknown > what i should test next i am learning that agents do not become useful because they have more tools they become useful when one small job is clear enough to repeat tomorrow: making the researcher better at finding useful AI discussions

the transcript was just a test input i gave it to the researcher i built yesterday it separated: > what the speaker actually said > claims that still need checking > missing context > what is worth testing myself the useful part was not another summary it was a path from an interesting claim to something i can verify one input one brief one next experiment

downloaded a youtube video i did not have time to watch fed it to my local agent instead ran it through whisper > full transcript > timestamps > no summary > no rewriting > just the words the point is not speed a video you cannot sit through becomes something you can skim, search, save, and come back to later reading scales in a way watching does not testing next whether these transcripts turn into research briefs without losing what mattered in them

most people talk about AI agents like they are a product category yesterday i tried to build one i installed ollama and connected it to hermes agent now the model runs locally on my own machine no API bill for every message getting an agent to answer is the easy part giving it a useful job, the right tools, memory, and limits you can trust that is the real work next i am testing whether it can handle: > research > summaries > repeated tasks without turning into another demo

the useful warning this week was not another model card it was this: the window we use to watch the model think is fogging chain of thought was the sensor then tools, other agents, and the model rewriting its own steps ate the edge of that sensor so "it explained itself" is no longer a review if nobody sits on the last send, you are trusting a thought trace that was never built to stay honest once the work left the chat

people are picking a winner between astra and fable that is the wrong seat one is better at driving the computer one is better at long messy work that split is real. it still does not finish the setup most people bolt both strengths onto one bot, then dump a skill pack on top and call it a stack the model is not the failure the unread file is split the job first then pick the model then decide who is allowed to reject a skill before it can send