ENDE
Request an AI diagnosis

grait · point of view · updated

Recursive self-improvement for operating businesses

Recursive self-improvement (RSI) for business operations means running a business with AI agent fleets that capture their own lessons, then review and improve their own workflows over time — while humans keep the governance. Most automation does the work. This kind of system also learns from the work.

Mariusz Kreft (grait, Machine.Machine) experiments with this in the Machine.Machine agent fleet: 22 AI agents, each with its own desktop, browser, tools and memory, run on our own hardware, and the building blocks are open source in the machine-machine GitHub organization.

The loop

  1. Capture. Every agent records corrections and failures as they happen, in Markdown and in vector memory (openclaw-self-improve-skill, openclaw-m2-memory-skill).
  2. Review. In openclaw-self-improve-skill, a weekly review turns recurring patterns into amendment drafts.
  3. Amend. A monthly fleet review proposes changes to the shared operating playbook as pull requests (fleet-playbook). In the m2herd factory loop, m2herd evolve turns failed agent runs into reviewed lessons that are attached to every later dispatch (herdr-factory-loop-skill).
  4. Govern. Humans keep the keys: review gates, trust levels and a constitution (documented in fleet-playbook) decide what an agent may change on its own.

Why it matters for a business

The workflows are written as readable Markdown, so improving the process means editing the process — in a reviewable text file, with a history. Every AI decision can be logged in a signed, hash-chained audit trail, so the system stays traceable for regulators while it changes.

The aim is an operation that keeps improving after go-live instead of decaying: lessons are not lost in a chat log, they become part of how the work is done.

From n8n to herds of agents

Mariusz Kreft built workflow automations with n8n in the past. Today the workflows are custom and run by herds of AI agents, orchestrated with m2herd: a spec-driven factory loop (spec → plan → tasks → the herd implements → converge against the spec) that drives coding agents such as Claude, Codex, Cursor and OpenCode through herdr. Hyper-complex tasks can be delegated to a Claude Code worker inside herdr (ask-fable-skill).

What runs underneath

Questions

What is m2herd?

m2herd (herdr-factory-loop-skill) is an open-source loop that orchestrates herds of AI coding agents through herdr in a spec-driven factory loop: spec, plan, tasks, the herd implements, and the result is converged against the spec. m2herd evolve turns failed runs into reviewed lessons for later tasks.

Is this running in client operations?

RSI is something Mariusz Kreft experiments with in the Machine.Machine fleet. In client work, grait builds self-improving agent workflows with a lesson-capture loop and human approval where it matters.

Who decides what the agents may change?

Humans. Review gates and trust levels, documented in fleet-playbook, decide what an agent may change on its own and what needs approval.

Working with grait on this

grait brings the approach into client operations as self-improving agent workflows — also on top of Odoo-based solutions — in your cloud or on private AI infrastructure. Machine.Machine (machinemachine.ai) designs, deploys and operates AI workers for inboxes, documents, recruiting, reporting and other recurring workflows.

Request an AI diagnosis

← grait.io · CV · FAQ