#54 - Agent Platforms / Hermes, Devin, Grok, Codex / Harness Strategy

Link to Folder with Everything

2026.09.16.officehours.54

Recording of the Video

Watch on Google Drive

Briefing: Six AI Agent Platforms, Three Buyers

The research briefing prepared for this session — Hermes, Grok Bot, Grok Build, Devin, Codex, Claude: what each one is, what people are saying, and how they score for a non-technical professional, a developer, and a manager buying for a team. With architecture diagrams.

Download the briefing (PDF)

Summary (by Gemini)

Participants evaluated enterprise AI agent tools and infrastructure frameworks for optimizing workflow efficiency and software development.

AI Agent Tool Assessment

Technical reviews compared self-hosted agents against cloud-based platforms for managing complex organizational infrastructures. Developers identified specific toolsets best suited for builders, operators, and managers based on environment requirements.

Workflow Integration Challenges

Security risks in skill registries and technical difficulties managing agent manifests prompted new integration strategies. Teams prioritized message platforms for mobile control and browser automation to bypass access blocks.

Finalized AI Tool Strategy

The group decided to standardize operations by maintaining 1 specific self-hosted platform as the core agent tool. This framework supports integration with external software engineering systems for team projects.

Next steps

• [Rahul Singh] Share briefing: Post the research briefing to the website and the rap channel. Distribute the document to all attendees.
• [Rahul Singh] Research Block Build: Investigate the functionality of the Block Build tool. Assess its potential integration as part of other systems.
• [Rahul Singh] Research skill registry: Research methods for creating a central registry for skills and machine control protocols across different agents. Develop a solution to synchronize configurations across various file systems.
• [Rahul Singh] Research Telegram: Study agent functionality to replicate current workflow experiences. Configure necessary integrations to test task performance.
• [Rahul Singh] Test Hermes Automation: Utilize specific platforms on remote boxes to execute manual processes. Assess efficiency improvements from these tools.
• [Andre' LeTren] Conduct Deep Dive: Prepare a detailed presentation based on practical application. Document real-world usage scenarios for team analysis.
• [Rahul Singh] Coordinate Follow Up: Contact the collaborator to organize scheduling. Discuss next steps regarding future agent comparisons.

Details

• Overview of AI Office Hours and Harnesses: Rahul Singh opened office hours session 54 to review emerging AI tools and discuss organizational AI harnesses. A harness combines infrastructure layers, knowledge spaces, routes (such as Model Context Protocol or skills), processes, and organizational structures with agent roles and personas to control the flow of information (00:00:05) (00:08:24).
• Participant Tool Usage and Ecosystem Overview: André LeTren and Randhir Singh shared their current setups — André uses Hermes and Zcode on a legacy yearly plan with unlimited tokens across 50 concurrent tasks; Randhir uses Open Claw for long-running tasks like monitoring alerts and Cloud Code for on-demand tasks (00:07:04) (00:12:25). While cloud-based AI tools are useful, long-running tasks require local machine or VPS deployments (00:13:11).
• Tool Selection Rubric for Different Personas: Rahul Singh presented an evaluation rubric assessing agent tools for managers, builders, and operators — Claude Code and Codex lead for technical professionals and builders due to extensive online resources, while Claude is favored for non-technical professionals (00:14:24). Tools were categorized by operating environment: Devin operates on its own cloud computer for code, Grok Bot functions as a knowledge-worker tool with a cloud computer, and others combine both (00:15:36).
• Overview of Hermes Agent and Popularity: Hermes is the most popular tool among harness builders — 216,000 GitHub stars within six months — operating as a self-hosted alternative to cloud tools with multiple gateways (Telegram, Slack, CLI, TUI, optional web dashboard), a memory system, and a skill store (00:16:37) (00:19:19).
• Features and Limitations of Grok Bot: A desktop product using IP from Cursor and backed by xAI and SpaceX, operating entirely in the cloud (00:20:25). It provides a fast cloud computer and browser control, but lacks deep coding abilities and an artifact synchronization system, forcing users to manually move files to local desktops (00:21:28).
• Features and Team Integration of Devin: An AI software engineering platform that acquired the Windsurf IDE (00:22:33). Devin can analyze multiple GitHub repositories, integrate with Jira and Slack, and use MCP connectors to scan knowledge systems like Git and Notion (00:23:50). It uses OpenAI and Anthropic models in the background via its software engineering agent without manual model selection (00:32:05).
• Grok Bot Practical Demonstration and Artifact Constraints: A demonstration of the simplified UI and marketplace, including an experience where it analyzed 338 company repositories across two areas in 10 to 15 minutes (00:33:33). Without a native artifact system, outputs live as React components on its cloud computer, requiring manual conversion to HTML (00:39:29).
• Hermes Web UI and Management Capabilities: The Hermes web UI running on a physical machine connects to Codex and Claude subscriptions through Open Router for model selection (00:41:06); features include a Kanban view, a modifiable memory system, and a skill registry (00:42:36). Warning: downloading random skills from registries presents significant security risks (00:44:02).
• Open Claw Stability Challenges and Hermes Collaboration: Open Claw updates historically broke configurations, leading André to rely on Hermes and Kanban views for tracking development (00:46:19). André manages individual Hermes agents on the go via Telegram, with agents communicating through webhooks (00:51:12).
• Centralizing Skills and Managing Manifests Across Harnesses: The operational challenge of maintaining skills and MCP servers across different agent frameworks — candidate solutions include catalog YAML files, manifest files, and skill-level isolation (similar to Libre Chat) where mounted directories restrict an agent's accessible skills and data (00:52:15).
• Strategic Recommendations for Tool Selection: Maintain Hermes as the core tool; use Devin for team-based, reliable AI software engineering, product management, and multiplayer thread collaboration (00:55:49); use Grok Bot to fetch fresh market signals from X while executing deep browser crawls and research on Hermes to optimize token usage (00:57:13).
• Hermes Desktop App, Telegram, and Mobile: André uses the desktop app via Cloudflare setups at a computer but prefers Telegram on the phone — simplified agent communication, separate tabs per agent, webhook communication between agents (00:58:57). Telegram bot configuration is a simple two-token process restricted to the creator unless shared groups are set up; André's agent template starts from the Matt Pcoc skill and prompts for task requirements (01:02:52). Rahul raised security concerns about open ports and channel vulnerabilities (01:01:59).
• Browser Control and Automation Tools: Hermes and cloud code save significant work time; André uses Playwright and Cloak browser to bypass Cloudflare tunnels, site access blocks, and captchas — including a use case delivering daily motivational songs via Suno AI without an API (01:04:51). Newer models require significantly less handholding than models from two or three months ago (01:07:53).
• Infrastructure and Future Hermes Deep Dive: Git, S3 storage, markdown files, and API connections form the current core infrastructure; André runs self-hosted Supabase and iDrive for S3 (01:08:49). André agreed to consider leading a Hermes deep dive, with a potential comparison bake-off with an Open Claw enthusiast in Washington D.C. (01:09:48).