> ## Content Index > Fetch the complete content index at: https://bayesiantheory.com/llms.txt > Use this file to discover other available public pages before exploring further. # My COO doesn’t sleep, eat, or need a salary - URL: https://bayesiantheory.com/my-coo-doesnt-sleep-eat-or-need-a/ - Published: 2026-03-23T11:39:26.000Z - Updated: 2026-10-05T14:45:24.000Z - Description: How harnesses made everything we know about agents irrelevant - Author: Gregg Bayes-Brown - Tags: Briefing, AI, AI enablement, Newsletter, #Migrated-1791211413550, #Import 2026-10-05 14:44 It’s 9 AM on a Monday morning. At 7:30, I had a check in with my COO. They came prepared with a full briefing: My emails had been reviewed, my to-dos addressed, and my calendar prepared. It’s been a busy hour and a half. Ideas from the weekend have been researched. Business development pitches have been drafted. Updates to builds I’m working on have been planned awaiting execution. My whole work week has been planned out: pomodoro sessions for the AM, deep dives in the PM. Stuff I couldn’t get around to last week resurfaced, ideas I’d forgotten I had given time for proper consideration, and my travel plans all booked. Not for me though. I lifted weights to Slayer and ate a sausage sandwich while drinking coffee. Welcome to 2026. [Subscribe now](#/portal/signup) ## We gave it hands Just six months ago, I was in a very different place. I was working at a FTSE 250 trying to figure out how to integrate AI agents into the business. If you’ve been in that position, you’ll know the feeling. Every meeting presents a new layer of organisational quagmire. Use cases surface governance questions. Governance introduces its own debates about accountability, security, and data management. Each conversation requires both technical and professional translation across multiple domains. The technical challenge of getting agents to work reliably was real. The frustration from building on Copilot Studio had me in anger management. And the organisational friction - such as any and all momentum getting crushed by the monolith that is IT – has left marks on my very soul. All of that friction assumed a specific architecture: autonomous agents replacing workstreams. Bots doing jobs. With it comes rational takes on ethics, governance, and ownership – all underpinned by multifaceted anxiety. When you hand a workstream to an autonomous agent, you genuinely do need to worry about all of that. Far from flicking a switch, AI adoption is a contact sport. Building an agentic system requires fundamentally redesigning a whole workflow. Often, these procedures aren’t clearly defined. When you do find a procedure, it’s normally one your system conflicts with, so several rounds of meetings are required. Nor are agentic systems truly autonomous: key inputs, triggers and reviews are required by humans for them to function. This requires training, securing buy-in, breaking of habitual working patterns, maintaining consistent engagement, and managing anxiety. It is little wonder that many companies have seen their AI adoption stall. It’s no simple task: it requires a fundamental top to bottom overhaul in how a company operates. For a tech which is prone to errors and breaking, that’s quite the ask. Then harnesses came along. ## Models, apps and harnesses Jensen Huang stood on stage at GTC last week and called OpenClaw, an open-source harness which took the AI world by storm at the start of the year, “the next ChatGPT.” By that point, it had already crossed 250,000 GitHub stars — more than the Linux kernel. Tencent integrated it into WeChat yesterday, meaning one billion people just got access. Meanwhile, my COO, chief orchestration officer, is Anthropic’s Claude Cowork. But what are harnesses? First, models. These are your GPT 5.4s, your Opus 4.6s, your Gemini 3.1s. 2025 agentic systems used API calls – essentially mailing them a pre-planned prompt – to build everything around them. The majority of people interact with models via apps: chatbots, wrappers, products with an LLM somewhere inside. These come with system prompts – this is why ChatGPT and Copilot feel so different despite being the same model – and app-specific tools, such as NanoBanana or CustomGPTs. A harness is an environment where an AI model has access to a persistent, extensible toolkit. Skills, plugins, MCP servers, memory, a file system - it’s operating within a context that equips it. The model stays the same – Opus 4.6 can be called via API, chatted with on Claude.ai, or operate the Claude Code harness – but the outputs vary wildly. If I said “plan my day” to an API, I’d get generic steps on how to plan the average day. Claude.ai will look at recent conversations and highlight anything that’s been top of mind. Claude Cowork spins up automatically at 0730, connects to my Gmail, calendar, and taskboard, and follows a set list of tasks to accomplish. If I haven’t done something, it moves things around. When it hits something ambiguous - a client relationship, a political decision, a judgment call that depends on context I haven’t written down - it asks me. Then it carries on. ## The seismic shift With harnesses, a company’s structure doesn’t need to change. Humans are still making decisions, handling the sticky political stuff, and reading subtext in meetings. The messy, ambiguous, context-dependent work that organisations are actually built on remains. The org stays the same. All those governance nightmares assumed you were replacing workstreams and workers. Harnesses don’t replace either. They equip the person already doing the work: individual power gets multiplied. The accountability, the collaboration, and the navigating an organisation all remain with the worker. You aren’t stripping a five-man team down to one with a few agents. You are turning a five-man team into a twenty-five-man team without increasing headcount. This shift leaves organisations with three choices right now. You can do nothing and watch competitors pull ahead. Remember those 5% who were succeeding at AI adoption from MIT’s report last year? According to BCG, there’s now a gap, and that gap is accelerating. You can invest in 2025-era agent infrastructure - custom orchestration, RPA with LLM duct tape, heavy governance and IT gatekeeping. Stuff which has already been made obsolete by a system built by one Austrian bloke in his spare time. Forty-two percent of companies abandoned most of their AI initiatives in 2025, up from 17% the year before. Eighty-eight percent of proof-of-concepts never reached production. 90% of CEOs said they saw no productivity from AI over a three-year period. The pattern is clear: adoption as we knew it – a job for IT – is failing. Or you can recognise that the organisation doesn’t need to reshape itself around AI. It needs to enable every employee to be the interface between the messy human reality of the business and intelligence that can execute. For that third option to work, everyone needs to be enabled. Not a Centre of Excellence with a backlog. Not a pilot with twelve people and a Jira board. Everyone across your organisation as a priority. Teach people how to use harnesses, or wait for the live demo when AI-native administrators come to close down your business. Now, if you’ll excuse me, my COO needs to build my website, so I’m off to the sauna.