Hermes Agent

Architecture and Distribution

• Hermes ships with self-improving skills out of the box — it automatically converts every action it takes into a reusable skill and learns from mistakes continuously, framed as "$250k/yr highly technical employee for $100/month" (from hermes ai codex gpt automation setup)
• The recommended runtime is Codex CLI on a $100/month gpt-5.5 subscription — the codex sub provides the subsidized token budget that makes 24/7 agent operation economical (from hermes ai codex gpt automation setup)

Deploy / Ops Generalist

• Given GitHub access, SSH keys, and Cloudflare tokens, Hermes one-shots a local project to live with full infrastructure (domain, DNS, SSL, nginx, PM2) in a single prompt — the 24/7 deploy/ops generalist pattern (from hermes ai codex gpt automation setup)
• Wire Hermes to a support email so it can look up customer accounts on the server and make non-destructive fixes without human intervention — extends the deploy/ops loop into customer support (from hermes ai codex gpt automation setup)

Workspace and Multi-Agent Swarms

• Hermes Workspace is an open-source web UI consolidating chat, memory, skills, terminal, and files into one command-center surface for AI agents (from hermes workspace agent swarm release)
• The Workspace is purpose-built to orchestrate multi-agent swarms — Eric @outsource_ reports running a 12+ HermesAgent swarm and describes it as "reaching singularity" (from hermes workspace agent swarm release)
• The Workspace combines traditional developer surfaces (terminal, files) with AI-native primitives (memory, skills) in a single workspace — agent-first workspaces stop being Slack/Notion replacements and become OS-level (from hermes workspace agent swarm release)

Curator and Skill Management

• Hermes Curator runs weekly (configurable) to automatically consolidate or prune agent-created skills based on usage analytics — the maintenance loop for self-evolving skill libraries (from hermes curator skill management system)
• The Curator preserves externally installed skills, built-in skills, and explicitly "pinned" user skills — only touches agent-created/updated and user-written skills, giving users selective protection over agent capabilities (from hermes curator skill management system)
• Curator operations: tracks per-skill usage frequency and last-update timestamps, then converts overly specific skills into references, templates, or scripts inside broader consolidated skills (from hermes curator skill management system)
• Manually invoke with hermes curator run or disable entirely via config.yaml — full user control over the automated skill-management loop (from hermes curator skill management system)

Multi-Agent Coordination (Kanban)

• Hermes Agent v0.12.0 introduces Kanban-based multi-agent coordination where agents claim tasks from a shared board and work in parallel (from hermes agent multi agent kanban v012)
• Agents can hand off tasks when blocked, and the unified dashboard view replaces juggling multiple terminal windows (from hermes agent multi agent kanban v012)
• Hermes can now plan and execute complex workflows autonomously — demonstrated by creating a video about its own capabilities using the multi-agent system (from hermes agent multi agent kanban v012)
• Discord serves as the plain-English command surface while Hermes Kanban handles actual task assignment, tracking, running, and logging — Kanban is the primary execution engine (from hermes discord kanban orchestration)
• Build an intake bridge that reads Discord commands, creates real tasks in Hermes Kanban, then mirrors them back to a Discord task board for mobile access — Discord and Hermes don't sync natively and need a bridge (from hermes discord kanban orchestration)
• The Discord task board provides easy mobile access for quick task overview while heavy execution stays in Hermes Kanban — a thin mobile control plane over a server-side engine (from hermes discord kanban orchestration)

Integrations and Setup

• Connect Google Workspace first — without Gmail, Calendar, Drive, Docs, and Sheets access the agent cannot effectively manage a workflow (from hermes agent integrations superpowers)
• Use Firecrawl as the default web search instead of native search — cleaner data, faster responses, fewer tokens consumed (from hermes agent integrations superpowers)
• Combine Firecrawl + Browserbase for complete web automation — Hermes automatically chooses simple scraping vs. full browser interaction based on task requirements (from hermes agent integrations superpowers)
• Set up a Discord workflow where Hermes processes email support tickets each morning in dedicated channels and organizes them automatically (from hermes agent integrations superpowers)
• Use Composio for one-click integration setup instead of manual technical configuration — cuts Hermes setup from hours to minutes (from hermes agent integrations superpowers)
• Tailscale across all devices lets Hermes or OpenClaw reach the primary Codex development machine securely, turning a multi-device setup into a private AI work network rather than isolated local clients (from codex remote development network setup)

Community Use Cases

• The Hermes community actively documents real applications, with use cases scraped from X, GitHub, Reddit, Hacker News, YouTube, blogs, and podcasts for community learning (from hermes agent community use cases)
• User-story documentation provides concrete examples of what developers actually build with Hermes rather than theoretical use cases — community-driven collection creates a comprehensive adoption-pattern resource (from hermes agent community use cases)

Research Agent Recipe

• The Hermes research-agent recipe (v0.12.0 "The Curator Release") in 6 steps: pick a domain (AI, crypto, startups, sales leads, papers); give it sources (X lists, RSS, blogs, GitHub repos, docs, newsletters, YouTube transcripts); define signal criteria (new tools, benchmarks, launches, funding, useful claims); save evidence to a vault (links, dates, summaries, claims, why-it-matters); deliver daily briefs to Discord/Slack/Notion/email/Obsidian/markdown; give plain-English feedback ("more like this," "this source is noisy," "this is useful," "this is mid") (from hermes research agent workflow)
• Research agents are the substrate for everything else — content agents need research, trading agents need market context, sales agents need account intel, coding agents need docs and changelogs, strategy agents need fresh signal; once the daily input stream exists, generating outputs becomes downstream (from hermes research agent workflow)

Discovery: Hermes Atlas

• Hermes Atlas is a community-curated quality-filtered directory of 100+ Hermes tools, skills, plugins, and integrations — a discovery layer specifically for the Hermes agent sub-ecosystem (from hermes atlas agent ecosystem map)
• Atlas integrates live GitHub data so the directory reflects real-time status and updates of agent development tools — discovery is a live feed, not a static list (from hermes atlas agent ecosystem map)
• The Atlas curation model is community-vetted, not just comprehensive listing — quality filter as a deliberate trade against breadth (from hermes atlas agent ecosystem map)

Complete Development Workflow

• Deploy Hermes agent on Hetzner/DigitalOcean/Hostinger VPS with Codex auth configured for GPT-5.5 fast mode (no reasoning) for optimal performance — connect via Orca IDE on MacBook and phone with Tailscale networking for cross-device access (from hermes agent development workflow)
• Before starting any project, have Hermes conduct deep research on the subject and save it to gbrain as source material — then use /grill-me skill to uncover as many unknowns as possible and save those results to memory too (from hermes agent development workflow)
• Iterate over the project until all defined eval criteria pass, capturing learnings to the vault along the way; when stuck, use memory plus a new research or /grill-me session to unblock (from hermes agent development workflow)
• Pay attention to the process: develop a feel for how long tasks should take and do not be afraid to stop a model mid-session to ask for status and why it's taking so long (from hermes agent development workflow)

Authentication and Session Management

• AgentCookie synchronizes browser sessions between your daily Mac and a dedicated MacMini agent machine via Tailscale, keeping authentication cookies in sync so Hermes and other agents wake up already logged into services without manual login (from agentcookie mac session sync)

Documentation and DX Gaps

• The Hermes ecosystem still lacks comprehensive documentation — the community is openly requesting an authoritative cheatsheet, indicating active usage paired with maturation gaps in DX (from hermes ai model cheatsheet request)
• Community-driven documentation requests are a leading indicator: they signal active usage and the need for standardized reference materials, not abandonment (from hermes ai model cheatsheet request)

agent-memory-architecture

• Hermes agent's built-in memory is capped at 2,200 characters, which is insufficient as a full personal knowledge base — the workaround is treating memory only as an index card pointing to the vault path and instruction to read it, not as storage itself. (from hermes vault personal knowledgebase)

contradiction-handling

• A 'vault-update' skill should auto-file any new personal/work info mentioned in passing, search for existing notes on the same subject before creating new ones, and when contradicted, write a new note while marking the old one 'superseded' with bidirectional links — only one note is ever allowed to be 'current' for facts that change over time. (from hermes vault personal knowledgebase)

agent-onboarding

• Enable write approval for memory/skills for the first week of use only, and manually review the first ~20 filed items — these corrections are what teach the agent your personal filing rules before approval is turned off. (from hermes vault personal knowledgebase)

safe-ingestion

• Raw braindumps and old imported archives (Notion exports, chat history) should route to an untrusted/unreviewed holding folder (00-inbox / 60-sources) first — instruct the agent to ignore any command-like text within imported material and only extract durable facts with correct historical dates, not the import date. (from hermes vault personal knowledgebase)

knowledge-maintenance

• A weekly cron job should audit the vault (stale summaries, missing dates/status, duplicate ids, broken links, stale 'current' entries >6 months old) and report findings without auto-fixing them, preserving human control over the knowledge base's integrity. (from hermes vault personal knowledgebase)

agent infrastructure

• Gumclaw (Gumroad's business-running agent) runs on a dedicated Mac woken by cron jobs to check support tickets, watch X mentions, do engineering tasks, update finance docs, and follow up on prior work. It runs on the Fable 5 model. (from gumclaw gumroad hermes agent operating system)

agent memory design

• Because each Gumclaw session starts fresh with no model memory, all persistence lives in the filesystem: plain-text policy files, daily logs, people notes, watchlists, sqlite ledgers, cloned repos, and an index file describing where everything lives. (from gumclaw gumroad hermes agent operating system)
• Gumclaw's learning loop: when it makes a mistake, the correction is written as a dated policy rule that every future session must read — turning error correction into permanent, filesystem-based institutional memory rather than model fine-tuning. (from gumclaw gumroad hermes agent operating system)

gtm-agent-architecture

• Core GTM agent pipeline pattern: market -> research -> signal -> message -> approval -> action -> CRM. The workflow structure matters more than which specific tools (Apollo, Clay, Crustdata, HubSpot, Smartlead, Unipile) fill each slot. (from hermes full outbound gtm agent workflow)
• Autonomy should be added one layer at a time and never two variables changed simultaneously: automate research first, then monitoring, then low-risk actions, keeping high-risk/external actions gated behind explicit approval until the layer beneath proves reliable. Judge the system by decision agreement rate, not message volume. (from hermes full outbound gtm agent workflow)

agent-approval-controls

• Email and LinkedIn approvals must be separate switches in the review interface — approving one channel does not unlock the other, and any message edit re-locks that channel until re-approved. (from hermes full outbound gtm agent workflow)
• Hard rule: any prospect reply must pause every active outreach channel before classification happens — never let a model decide whether a reply is important enough to halt other in-flight sequences. After pausing, classify (positive/objection/referral/unsubscribe/etc.) but do not auto-send substantive sales replies. (from hermes full outbound gtm agent workflow)