the fundamentals of building Hermes bots (marketing edition) bot mode is how you can easily run an agent marketing team inside hermes desktop. one bot per vertical, each with its own harness and the ability to send work to each other but like any other agent, it still needs the same building blocks: context, harness, and loops here's how each one block views up when you build a hermes bot: 1. context context is what data and information you feed the agent. the model only works from the context window (the text it can see for one run), it remembers nothing from the last run you pull that from a knwoledge layer (ex company brain). the brain holds past campaigns, strategy docs, offers, brand rules, voice rules, transcripts, and examples of what good looks like, a research bot can also pull external context, competitors, channels, and trends a skill then decides which pages enter the context window for the current job. the content skill pulls in voice rules and examples, SEO skill pulls in the brief, SERP notes (search results page notes), and previous ranked content, campaign skill pulls in the offer, the ICP (ideal customer profile), and the last cohort's numbers. without the good context, you will just get sloppy generic output 2. harness the harness is what lets the bot load a skill, call a tool, keep a file, and stop for your sign-off. models alone only generate text. in bot mode that harness is the bot itself the harness pieces: > model: cheap and fast for throughput, stronger for review and taste work > skills: the procedures it loads for a job, steps, files to read, tools to call, what done looks like > tools: what it can do in general, browser, file read and write, web search, local scripts > connectors: hooks into your stack, CRM, CMS, search console, ad accounts, email, analytics > memory: what it keeps between jobs, notes, progress logs, past outputs for that vertical > role + voice rules: identity, what outcome it owns, what it never does > control: timeouts, token budgets, sign-off points, what it can spend or send without you example: the content bot can draft and queue. it cannot post live, send the sequence, edit the CRM, or move spend without a sign-off 3. loop/s how the bot keeps working until the output passes the quality we want (the pass/fail checklist for the output). it drafts, a check scores the output against that rubric, and any miss goes back with the failing line named. you set the goal, the check, and the budget before it runs content loops on voice and anti-slop. SEO loops on the writing brief and search position. a campaign loop tests variations, reads performance, adjusts creative.. loops also run as scheduled jobs, cron jobs that fire on a timer. daily SEO check. weekly content batch. nightly competitor scrape into the research bot. same loop, cron is the start signal build from context out: 1, put company context in the brain 2, create one bot per vertical and attach its harness 3, add a loop with a clear done state, including cron where the job should run without you starting it