In a small, practical demonstration of what many technologists describe as the next wave of automation, a Google senior product manager, Shubham Saboo, has published a how‑to for assembling a six‑agent AI workforce that runs unattended and saves him several hours of work each day. Using the open‑source agent framework OpenClaw, a single Mac Mini and a mix of commercial models, Saboo built an automated pipeline that scans the web, drafts social posts, reviews code and prepares newsletters while he sleeps.
Saboo’s setup deliberately rejects the idea of one omnipotent model and instead partitions tasks among specialised agents — each given a distinct personality and remit. He named them after sitcom characters to encode behavioural expectations: ‘Dwight’ does research, ‘Kelly’ drafts X posts, ‘Rachel’ writes LinkedIn pieces, ‘Ross’ handles engineering reviews, ‘Pam’ edits newsletters, and ‘Monica’ coordinates and enforces scheduling. Communication between agents is not achieved through complex APIs or message queues but by writing and reading Markdown and JSON files on disk, a design Saboo argues is more robust and easier to debug than real‑time orchestration.
The system combines simple infrastructure choices with practical governance. Agents run on an isolated machine with separate credentials, interact with Saboo via Telegram, and follow cron schedules backed by a heartbeat monitor that detects missed runs and triggers retries. Saboo also implemented a two‑tier memory scheme: daily logs archived as records of each session and distilled long‑term memories that capture recurring corrections and style preferences. That feedback loop is central to the approach: incremental, human‑in‑the‑loop refinements turn mediocre initial outputs into progressively reliable behaviour.
Costs are modest by tech‑industry standards. Saboo reports running the six‑agent “team” for under $400 per month: a Mac Mini as the host, OpenClaw as the orchestration layer, and subscriptions / API usage across models such as Claude’s Max plan, Google Gemini, and specialised speech or utility models. He estimates a daily saving of four to five hours — a meaningful productivity gain if replicated at scale — and argues the strategic value lies not in any single model but in the evolving system of memory, prompts and scheduling that accumulates organisational knowledge.
The experiment has landed in a broader debate about how AI productivity tools will diffuse. OpenClaw has drawn attention for its rapid adoption in developer communities, prompting public comments from figures such as Nvidia’s Jensen Huang about how fast such frameworks are spreading. Saboo’s write‑up, posted on X and GitHub alongside a popular collection of LLM tools, has become a template for others asking not just how to run an agent, but how to build a resilient, evolving agent ecosystem.
For managers and organisations, the lessons are practical: break work down into specialised, bounded tasks; use simple, file‑based interfaces for information exchange; keep models sandboxed; and invest effort in explicit, human feedback loops rather than hoping a single model will generalise across roles. For policymakers and security teams, the example signals new vectors for risk: automated workflows that access external APIs, scale cheaply and embed long‑running memories raise questions about data leakage, credential management and responsibility when agents act at scale.
This is not a turnkey solution to broader labour or governance questions. The agents Saboo describes still require careful human oversight, design and curation; they make mistakes that must be corrected through repeated feedback cycles. Yet the experiment illustrates how decentralised, open frameworks can rapidly lower the cost of building persistent automation — a pattern that will accelerate both adoption and the difficulty of regulating the technology effectively.
As firms and individuals replicate the model, organisations will face trade‑offs between agility and control. OpenClaw and the mix‑and‑match approach to models and channels (Telegram, cloud APIs, hosted models) accelerate innovation, but they also shift the locus of control to whoever configures the memory, credentials and scheduling. That makes operational hygiene — credential isolation, audit trails and explicit escalation paths — the decisive factor in whether these systems increase productivity safely or amplify risk.