🤚 The Open-Palm Illumination

There was a time — and by “a time” we mean February — when the act of coding with AI involved one human, one terminal, and one agent doing its level best not to hallucinate your database schema into oblivion. Those were simpler days. Quaint days. Days we will tell our grandchildren about while they stare at us with the vacant pity reserved for people who once drove their own cars.

Enter Alex Finn, who has apparently decided that one AI agent is for amateurs, and has unveiled a workflow so aggressively automated that the developer’s primary role is now sending a Telegram message and going for a walk.

The setup: OpenAI’s Codex builds the code. Anthropic’s Claude Code reviews it. NousResearch’s Hermes Agent orchestrates the entire affair from a Kanban board like a project manager who actually delivers. The user sends a single message — a goal, not a task, not a ticket, not a Jira epic with fourteen sub-stories — and the system handles the rest. Planning, implementation, code review, iteration, and board management. All of it. Autonomously.

Your standing desk just became a fainting couch.

👐 The Two-Handed Reality Check

Let us discuss what is actually happening here, because the implications are the kind that make engineering managers refresh LinkedIn.

At the heart of this unholy trinity sits Hermes Agent’s Kanban board — a durable task-tracking system backed by local SQLite, shipped in Hermes v0.12 and v0.13 across April and May 2026. This is not a pretty Trello clone. This is a multi-agent coordination layer with heartbeats, zombie detection, retry budgets, and a hallucination gate that prevents agents from confidently shipping nonsense to production. (A feature that, frankly, many human engineers could also benefit from.)

The magic ingredient is the /goal primitive — a first-class command that locks an agent onto an objective across turns. After every turn, a judge model reads the agent’s output and decides whether the goal has been met. If not, the agent keeps working. If yes, the card moves across the board. Both Codex and Claude Code now support this primitive, which means Hermes can treat them as interchangeable workers: assign tasks to whichever agent is best suited, swap them without workflow disruption, and maintain a single source of truth on a unified board.

The workflow, in practice:

• User sends a development goal via Telegram to Hermes
• Hermes creates a Kanban card and selects the appropriate agent
• Codex implements the feature, writing and testing code autonomously
• The card progresses; Claude Code picks it up for review
• Claude provides feedback, requests refinements, and the cycle continues
• The user monitors everything from a single board — remotely, passively, possibly from a beach

This is not theoretical. This is not a demo. This is a man building production software by texting a chatbot from his phone and watching colored cards slide across a screen.

🌿 The Gentle Awakening

We should pause to acknowledge the philosophical shift embedded in this workflow, because it is significant and it is weird.

The 2025 paradigm was: “I use an AI tool to help me code.” One agent, one human, collaborative pair programming. Intimate. Personal. You and your Claude, against the world.

The 2026 paradigm is rapidly becoming: “I manage a team of AI agents who code for me.” The human has been promoted — not out of the loop, but above it. You are no longer the developer. You are the engineering director who sets objectives and reviews outcomes. Your tools have tools. Your agents have agents.

As Alex Finn himself noted on X: “There’s this crazy tribalism in AI where people always feel the need to take a side — Claude Code or Codex, OpenClaw or Hermes. You’re missing out on 90% of the power if you do this.” The correct answer, apparently, is all of them, simultaneously, coordinated by yet another AI.

This is also, incidentally, the Agent Kanban thesis writ large — an open-source orchestration board that now supports Claude Code, Codex, Gemini CLI, GitHub Copilot, and Hermes as interchangeable workers on a single planning surface. The industry has moved from “which agent?” to “how many agents, and who’s managing them?” in roughly the time it takes to onboard a junior developer.

👑 The Crown Verdict

We are witnessing the emergence of the AI coding middle manager — and unlike its human counterpart, this one actually increases productivity.

Hermes Agent’s multi-agent Kanban, combined with the /goal primitive that makes different AI agents swappable commodities, represents a fundamental architecture shift. The coding agent is no longer the product. The orchestration layer is the product. The agents are labor. The board is the brain. The human is the board member who sets quarterly objectives and occasionally checks Slack.

NousResearch, with 765 contributors on Hermes Agent, has quietly built something that looks less like a developer tool and more like an autonomous software factory. Anthropic’s Claude Code and OpenAI’s Codex are now specialized workers in someone else’s pipeline. Whether either company intended this particular future is a question best left to their respective strategy teams and, presumably, their therapists.

The question is no longer whether AI can write code. It can. It has been able to for a while. The question is whether AI can manage AI that writes code. And the answer, delivered via a Telegram message and a Kanban board with zombie detection, appears to be: yes, and it doesn’t even need a standup.

Inspired by Codex Builds. Claude Reviews. Hermes Agent Runs It. by Alex Finn.

“We gave the robots a Kanban board, zombie detection, and a hallucination gate. They immediately outperformed three sprints’ worth of human output. The standup has been cancelled indefinitely.” — The Slap of Wisdom Automation Bureau, typing this from a terminal that no longer needs us