Alex Finn’s Hermes Agent Advice

LowLow — mostly Claude/Fable, but includes agent workflow patterns and a brief Hermes/OpenClaw teaser.

1. 01:45Use loops/autonomy to move from checking whether an agent did work to checking whether it is doing the right work.
2. 02:08Treat stronger models as thought partners: ask how they would approach or change the plan before building.
3. 05:40Use an advanced planning/interview mode before building complex productivity systems.
4. 07:45Paste the generated spec into a long-running goal-style workflow so the agent loops until success criteria are met.
5. 11:19Adjacent workflow: loop over Linear tasks periodically so agents can pick up new work automatically.

HighHigh — direct Hermes Desktop walkthrough.

1. 00:54Use Desktop sessions/threads for each area of life and pin important ones so scheduled sessions do not bury them.
2. 03:07Use Artifacts as a second brain for links, files, images, and generated outputs.
3. 04:43Disable unused skills/toolsets to reduce context, tokens, noise, and cost.
4. 06:13Use the Cron UI to verify, manually create, test, and pause scheduled jobs.
5. 07:27Use profiles only when roles need different memories, skills, personalities, tools, or models; otherwise use subagents.
6. 10:22Store API keys in the settings/API-key UI instead of pasting secrets into chat logs.

LowLow — mostly Claude/Claude Code; includes model-switching caution for Hermes/OpenClaw.

1. 05:29Switch Claude Code tasks to a new model first, but do not immediately force Hermes/OpenClaw onto it.
2. 06:23Wait for official Hermes/OpenClaw support before changing the agent model, because brand-new model IDs can crash agent setups.
3. 07:37Stay focused while agents work; do not prompt an agent and then waste the recovered time.
4. 09:28Use remote-control/mobile monitoring for coding agents when away from the desktop.

HighHigh — full Hermes setup/tutorial.

1. 01:51Think of Hermes as a 24/7 AI employee that learns goals, workflows, and preferences.
2. 08:33Install from the Hermes website command, then choose model and messaging channel; Alex favors Telegram.
3. 09:28Model tradeoff: Claude/Anthropic best but expensive; ChatGPT subscription is a moderate option; local/open models can be cheaper but less capable.
4. 13:44First task: brain-dump who you are, what you are working on, and goals so Hermes can save useful memories.
5. 15:23Schedule a nightly 2am proactive cron to build a micro-app/UI/system that advances your goals.
6. 21:02Use Kanban triage: add goals/tasks, let Hermes split subtasks, move them to to-do, and assign subagents.
7. 29:41Memories and skills are local markdown files; Hermes can review execution and improve future repetitions.

HighHigh — direct Hermes use cases.

1. 00:49Use /goal for long-running tasks, but meta-prompt first so the goal is detailed rather than vague.
2. 03:23Morning Kanban workflow: put AI-doable tasks into triage while you do human-only work.
3. 06:29Use browser/computer control for competitor app research: inspect site, console, features, and stack, then produce a report.
4. 08:58Build a memory wiki with subjects and daily logs so you and Hermes can review past work.
5. 10:36Use Tailscale to let Hermes administer/test across trusted devices.
6. 12:51Schedule a morning priority prompt: ask the #1 priority, propose support tasks, and update memories.

HighHigh — direct Hermes feature update.

1. 00:37Improved session recall lets Hermes retrieve prior sessions by date/time/topic without bloating current context.
2. 02:21/background lets Hermes run tasks in parallel while remaining available for normal chat.
3. 05:22Native Codex CLI support lets Hermes spin up Codex as a coding worker.
4. 07:02Computer use lets Hermes inspect and control apps remotely, e.g. calendar or Notion workflows.
5. 10:43Auto-Kanban task generation breaks triage goals into subtasks assigned to agents/subagents.

MediumMedium — coding-agent board workflow transferable to Hermes/Kanban.

1. 00:35Use a task board as an external second brain for coding agents.
2. 05:36Before coding, have the agent create detailed issues/projects for the app.
3. 09:02Acceptance criteria, priorities, and dates reduce drift in long coding runs.
4. 14:56Multiple coding agents can coordinate through the same board rather than manual context syncing.
5. 16:46Put durable process rules in AGENTS.md/CLAUDE.md: read issue specs, meet acceptance criteria, avoid unrelated refactors.

HighHigh — direct Hermes local-model setup.

1. 01:27Local models make 24/7 Hermes work more private and effectively free after hardware/electricity.
2. 05:40Use cloud-powered Hermes on the main computer to configure the headless/local model box.
3. 06:58Install Tailscale so Hermes can control the local machine over a private network.
4. 09:55Prompt Hermes to install a local model such as Qwen and validate it with a simple chat frontend.
5. 13:37Create a separate Hermes profile connected to the local model.
6. 16:25Use local Hermes for recurring research, transcript repurposing, self-improvement from videos, and lightweight coding.

HighHigh — direct Hermes setup/use cases.

1. 02:42Hermes self-improves by creating/updating skills when it encounters new or repeated tasks.
2. 04:31Hermes supports swarms/profiles; ask it to build a new profile for specialized agents.
3. 05:05Setup flow: terminal command, Telegram channel, and model choice; Opus best if budget allows.
4. 10:03Brain-dump career, goals, interests, and workflows into Hermes after setup.
5. 11:28Reverse-prompt Hermes for workflows based on what it knows about you.
6. 13:08Starter cron: daily AI news plus new AI tool recommendations based on your workflows/memory.

HighHigh — direct Hermes dashboard/Kanban guide.

1. 03:34Open the Hermes dashboard with `hermes dashboard`; use Kanban for many parallel tasks.
2. 04:44Kanban flow: triage → memory enrichment → ready → assigned execution.
3. 08:08Create a cron that checks Kanban every 10 minutes, expands triage tasks, and executes ready tasks.
4. 08:54Use a second admin/librarian profile for recurring Kanban management so the main agent is not blocked.
5. 13:35Profiles can have separate memories, skills, env vars, and roles; baseline: main, coding, research, admin.
6. 16:21Lower compression threshold to about 0.5 and use curator reports to reduce memory/skill bloat.

MediumMedium — coding-agent workflow transferable to Hermes.

1. 02:11Generate UI options before coding so design direction is clearer.
2. 06:31Run multiple sessions/agents in parallel for app, marketing, research, or other tracks.
3. 07:16Install specialized skills/plugins such as browser use, computer use, and video/design capabilities.
4. 12:17Use browser/annotation feedback and have the agent test itself.
5. 19:03Schedule recurring code-quality checks over recent commits.

HighHigh — direct Hermes + ChatGPT setup/use cases.

1. 01:01Hermes can act as a 24/7 computer employee: build apps, send email, check calendars, move files, and act on the computer.
2. 01:23Self-improvement through skills is a core advantage; repeated usage makes future tasks better.
3. 03:53ChatGPT 5.5 via OAuth is positioned as a lower-cost/high-limit Hermes brain versus Claude Opus.
4. 04:57Use Telegram BotFather setup to communicate with Hermes from multiple devices.
5. 06:35Use mobile Telegram prototyping and /steer to correct an active task mid-run.
6. 10:42Use cron for daily research or checking another agent every few hours.

MediumMedium — coding-agent practices transferable to Hermes.

1. 04:20Organize by project and run backend/frontend/planning sessions in parallel.
2. 05:33Enable computer use so the agent can test apps and interact with the browser/computer.
3. 08:10Start with image-generated interface options before implementation.
4. 13:43Ask the agent to click through and test all app features while you do other work.
5. 16:19Connect a project-management layer so multiple agents share issues and details.

MediumMedium — design workflow transferable to Hermes coding/design tasks.

1. 02:43Separate cheap visual exploration from high-fidelity design/prototyping.
2. 03:00Start with inspiration images/logos/styles and ask for several design directions.
3. 13:05Generate a brand guide/prompt after choosing the design direction.
4. 15:14Build a reusable design system before coding full interfaces.
5. 20:00Hand the design system/prototype to the coding agent for implementation.

HighHigh — OpenClaw tutorial with direct Hermes comparison and transferable setup workflows.

1. 13:53Use messaging surfaces intentionally: Telegram for daily chat; group topics for separate contexts.
2. 19:49Brain-dump identity, business, goals, and workflows to personalize the agent.
3. 20:36List manual computer tasks for a day, then ask the agent which workflows it can automate.
4. 23:00Use a project board/Linear-style workflow to assign tasks, branch, and PR code.
5. 30:50Hermes is described as faster/more token-efficient, while OpenClaw memory was stronger in his experience; use both if helpful.
6. 32:42Two-agent reliability: if one agent breaks, the other can inspect configs/code and fix it.

MediumMedium — coding-agent operating practices transferable to Hermes background/subagent work.

1. 01:15Use higher-autonomy modes to reduce repeated permission prompts when safe.
2. 02:40Front-load full task information at the start of a session.
3. 04:43Use Hermes/OpenClaw-style agents for on-the-go prototypes and tooling; use coding agents for deep production coding.
4. 05:56Ask for recap after long-running sessions so you can review unattended work.
5. 08:04Turn on notifications/hooks so long work pings you when complete.

MediumMedium — desktop coding-agent workflow transferable to Hermes Desktop/Kanban supervision.

1. 00:19Project-oriented desktop UI helps supervise many sessions with attention indicators.
2. 02:26Keep multiple local coding sessions on separate app areas to avoid conflicts.
3. 06:57Use accept-edits mode for speed, plan mode for reasoning/approval.
4. 10:11Use routines/scheduled tasks for nightly code review and bug fixing.
5. 15:02Tool split: coding agent for production app coding; Hermes/OpenClaw for prototypes, mission-control apps, and agent tooling.

HighHigh — direct OpenClaw + Hermes multi-agent setup.

1. 00:31Use two agents for reliability: when one breaks, the other can diagnose and fix it.
2. 02:28Run Hermes as a lighter/faster assistant or monitor while another agent handles heavier main work.
3. 03:53Allocate models by role: strong expensive model for main/orchestrator, cheaper model for monitoring/execution.
4. 07:57Planner/builder/reviewer workflow: strong agent plans, Hermes executes, strong agent reviews.
5. 12:11Schedule Hermes cron jobs to monitor systems built by another agent and alert on failures.
6. 13:44Use shared memory folders so agents have private logs plus shared decisions, mistakes, plans, and context.

HighHigh — memory/wiki workflow explicitly useful for Hermes too.

1. 00:00An Obsidian-backed memory system can improve both OpenClaw and Hermes memory.
2. 00:26Use daily logs that capture discussions, tasks, and work done.
3. 00:58Keep a mistakes file so corrections become future behavior improvements.
4. 01:18Use an agent-shared workspace for multi-agent context handoff.
5. 03:56Keep large memory in files checked on demand, not injected into every prompt.
6. 09:04Verify the agent is writing memories; if not, add the memory rules to AGENTS.md/agent instructions.

MediumMedium — model-routing/cost strategy transferable to Hermes.

1. 01:24Use a brain-and-muscle architecture: strong orchestrator plus cheaper executors.
2. 03:53For average hardware, keep a strong cloud orchestrator and route coding/other tasks to cheaper cloud/subscription models.
3. 07:53For strong hardware, use local models for coding, scraping, and data collection.
4. 13:56The orchestrator chooses the right cheaper muscles to preserve output quality while reducing cost.
5. 14:36Evaluate agent spend as ROI on 24/7 digital labor, not as a normal entertainment subscription.

MediumMedium — future coding-agent patterns transferable to Hermes expectations.

1. 00:55Proactive mode idea: agent proposes and builds follow-on tasks rather than only executing commands.
2. 03:10Dream/overnight mode: agent thinks about recent work and what to build next while you sleep.
3. 04:11Smart auto approvals sit between YOLO and constant manual prompts.
4. 05:26Risky approvals can be routed to messaging apps for user confirmation.
5. 05:53Agentic payments could remove blockers like buying hosting/accounts/templates, but would need explicit spending controls.

HighHigh — direct Hermes guide/comparison.

1. 01:52Hermes is described as lightweight and performant compared with OpenClaw.
2. 02:21Hermes can perform a complex briefing workflow and capture it as a reusable skill.
3. 03:21Hermes is transparent about tool calls, websites visited, snapshots, and what it is doing.
4. 03:42Hermes creates a Hacker News daily AI briefing skill, updates a cron job, and delivers via Telegram.
5. 07:22Obsidian-style memory improved both Hermes and OpenClaw memory.
6. 08:45Use Hermes and OpenClaw side by side; one can spin up or fix the other.
7. 12:38Model choices: Claude most agentic, ChatGPT cheaper/strong, local models possible for advanced users.

MediumMedium — local-model strategy transferable to Hermes local/open-model profiles.

1. 01:07Cloud models are expensive, less private, and hard to scale for 24/7 agents.
2. 03:19Local models are private, customizable, offline-capable, and effectively free after hardware.
3. 04:52Local models unlock 24/7 scraping, writing, coding, and recurring work without token bills.
4. 11:31Choose model size based on hardware; mentions Qwen, Nemotron, and MiniMax as options.
5. 13:19Ask the agent to inspect hardware and Hugging Face before recommending local models.
6. 15:30Use cloud orchestrator + local worker muscles for coding, research, scraping, and writing.

MediumMedium — OpenClaw use cases map to Hermes cron/memory/workflow patterns.

1. 00:23Daily memory tracker records discussions, coding, shipped work, and projects.
2. 02:18Prompt pattern: summarize every discussion into daily logs and refer back for project/day questions.
3. 03:22Trending content alerts can run through Discord channels for rapid response.
4. 05:54Use agents to vibe-code internal microapps/tools; use dedicated coding tools for serious consumer apps.
5. 07:11Reverse-prompt the agent for app ideas based on your workflows and memory.
6. 08:28R&D lab workflow: multiple models review recent work, debate ideas, and produce a memo.
7. 12:31Overnight employee cron: review goals and do one task that moves them forward.

MediumMedium — comprehensive agent-workflow tutorial transferable to Hermes.

1. 02:37Agent framing: give goals and get results; agent iterates and self-improves.
2. 14:18Use Telegram for daily chat and Discord for structured multi-channel workflows.
3. 19:23Schedule recurring research briefs via cron.
4. 24:06Brain dump goals/preferences/workflows, then ask for custom automations.
5. 25:44Use Discord as second brain with channels for briefs, drafts, alerts, and digests.
6. 28:12Do not blindly install third-party skills; have the agent inspect and recreate safe versions.
7. 32:47Build Mission Control for calendar, memory, docs, office/team views, and custom tools.

MediumMedium — subagent/org design transferable to Hermes multi-agent operations.

1. 01:22Start with one subagent rather than many; scale only after a clear use case.
2. 02:42Choose the first subagent based on daily work: coding, writing, or research.
3. 03:21Route tasks to the right model: strong orchestrator, coding model, cheap/local research/writing models.
4. 05:54Let the orchestrator configure subagents instead of manually editing configs.
5. 07:02Subagents enable parallel work, lower cost, and separate context windows.
6. 08:21Maintain an org/team screen listing each agent, model, and responsibility.
7. 15:21Reverse-prompt for org design: ask which subagents should exist based on your goals/workflows.