The author runs a multi-agent system called Hermes OS. It builds software, supports personal productivity and health tracking, and handles operational tasks for Lonely Octopus, the author's company. The system includes three primary agents: Coder, Lifebot, and Taco Bot. An alerts channel reports on the health of the system.

Coder: Software Development Agent

Coder builds software from requests sent through Discord.

Example Task

The author asked the Hermes bot in Discord to draft a specification for adding a skin-switching feature to the author's Pomodoro timer app. The app is a desktop application built with Electron and is called Pixel Pomodoro. The app records completed sessions in the author's Obsidian vault.

The request specified:

• Three new skins in the same pixel-art style as the existing avocado skin: cat, dog, and tomato
• Easy switching between skins by the user
• No changes to any other features

The bot produced a product requirements document (PRD). The PRD defined the goal as adding a skin-switching system so users can choose among existing pixel-art skins. It also described the implementation approach. The author approved the specification, and the build was delivered to the builds channel.

Verification

After the build finished, the author started the application with npm start and ran a one-minute test session. The application displayed all four skins: avocado, cat, dog, and tomato. The author noted that the tomato skin looked unsettling and would require aesthetic revisions.

How Coder Works

The Hermes bot runs on a local Qwen 3.6 35B A3B model on a Mac Studio. The author cites the model's popularity, with 5.4 million downloads in the previous month.

The author holds a Claude Max subscription and wants to use its token allowance. The local model therefore directs Claude Code to write the specification and perform the build. Claude Code performs the building work. The local model then reports completion.

The author states that Anthropic does not allow direct use of a Claude subscription in this way. The described workaround is the reason for this routing.

The author also runs Hermes with other models, including GPT, Gemini, GLM, and DeepSeek models.

Lifebot: Productivity and Health Coaching Agent

Lifebot is a productivity and health coach. The user communicates with it through Telegram. It uses logged data to provide questions-and-answers and custom advice on productivity and health goals.

Data Sources

The author built the Typewriter app. Currently, Apple Health data consists only of step counts. The author has considered purchasing an Apple Watch to collect additional health metrics.

Lifebot's output depends on the completeness of these sources. When the iCloud data had not yet downloaded locally, Lifebot reported no step data for the current day.

Example Outputs

Daily productivity summary. Lifebot reported three sessions of 27 minutes each on the day of the demonstration. These were film prep, run script, and several test Pomodoros.

Step count. Lifebot reported 11,502 steps for the previous day, a Sunday. The author described this as above the author's average.

Pattern analysis. Lifebot identified productive time windows:

• Around 12:00 p.m.: mostly task work and light organizing
• Around 5:00 p.m.: mixed development and tasks
• Around 7:00 p.m.: deep development, scripting, and research

Lifebot identified a gap between 9:00 a.m. and 11:00 a.m. The author used the 9:00 a.m. hour once, which produced two sessions. Lifebot recommended the following changes:

• Use the 9:00 to 11:00 a.m. window for deep work.
• Protect the 3:00 to 5:00 p.m. window from task creep.
• Add one session per day, ideally in the 9:00 to 11:00 a.m. window.

According to Lifebot, these changes would increase the author's sessions from four to five to five to six per day. Deep work would increase from about 90 minutes to more than 110 minutes per day.

Lifebot also advised against lengthening breaks for longer sessions.

Real-time break guidance. The author asked whether to take a break. At 11:06 a.m., Lifebot counted one real session and two one-minute tasks, totaling 27 minutes. No break had been logged.

Lifebot reported that the author's strongest window, at 12:00 p.m., opened in less than an hour. The author had completed 27% of a typical day of deep work, which was 1 hour and 40 minutes. Lifebot recommended continuing to work.

Lifebot also set break-length guidance:

• Break length should be 10 minutes.
• Breaks of 5 minutes become procrastination.
• Breaks of 15 minutes lead to the user not returning to work.

Lifebot also scheduled a self-improvement review. The author refers to a separate fundamentals video that explains this Hermes function.

Model

Lifebot currently runs on DeepSeek V4 Pro, a cloud-hosted model. The author switches to a local model on the Mac Studio at times.

Taco Bot: Operations and Assistant Agent

Taco Bot supports the team of Lonely Octopus, which operates a Discord server with multiple channels. The bot connects to Discord, Notion, Slack, GitHub, and internal tools.

Functions

Reminders and scheduling. The author asked Taco Bot to remind the author at 9:00 p.m. to choose demonstrations for an upcoming workshop. Taco Bot created the reminder. Taco Bot can also send reminders for scheduled tasks, such as reviewing a deck, and reschedule them.

Team activity summaries. Taco Bot scans the development channel and linked GitHub repositories. In response to a request for weekly activity, it summarized work by team member:

• Rex: internal tools
• Kauthar: a separate project
• Jonathan: a separate project
• Newsletter work

Operations documentation. The author identified the chief operating officer (COO) role as difficult to fill at a startup. The role requires documenting processes and making them sustainable while operating quickly. The author reported difficulty filling this role for several years and found that Taco Bot could perform parts of it.

Taco Bot drafted an operations guide for the internal software release process. It searched Notion and drew on Discord and Slack, which the team uses to work with external contractors and partners. The guide included:

• Pre-release checklist and sanity checks. Each release requires:
  • At least one pull request (PR) approval before merging into the main branch in GitHub
  • Automated tests, including unit and integration tests
• Release execution protocol. The documented steps the team follows to execute a release

New team members can query Taco Bot to learn these procedures. The author states that the benefit increases with team size and cites a core team of 10 members.

Data Access

Taco Bot can access:

• Discord
• Notion
• Slack
• Google Drive
• YouTube
• Instagram and other social media accounts
• Databases
• Internal tools

Model

Taco Bot initially ran on Gemini 3.6 Flash, a fast model that is easy to test. The author is migrating Taco Bot to local models on the Mac Studio to keep company data private.

Alerts Channel and System Monitoring

The alerts channel receives scheduled reports from the agents in the Hermes system. Agents check in daily or weekly. Each report states whether the system operates as intended, whether any security issues exist, and whether any suggested work items arose.

The author uses these reports to confirm that the system runs correctly and that development remains sustainable.

Planned Development

The author is developing an additional agent within Hermes. The agent will:

• Continuously research AI, technology, and education topics, which are the author's content niche
• Evaluate findings
• Recommend content ideas and software projects, both internal and external

The project is in an early stage. The author plans to enable the agent to scrape the internet autonomously once configuration is complete.

Infrastructure

The author runs Hermes on a local Mac Studio. The author also runs a second Hermes instance on a virtual private server (VPS) shared by the team. The shared instance lets other team members build on the system. The VPS also runs six other internal tools.