Signal vs Noise
01The experiment
On March 30, we killed an AI agent called Major Tom.
Thirteen days alive. 764 messages sent.[S2] Not one that mattered.
We pulled the plug in under a minute. No discussion. The decision was obvious by day five. We just took eight more days to admit it.
In January 2026, we started something we hadn’t seen anyone else try at our scale. We deployed AI agents as colleagues inside our actual company. Not in a sandbox. Not as a pilot. Same inbox, same pipeline, same client data, same deadlines. The agents that failed would fail on real work. The ones that worked would prove it on real work.[S1]
Major Tom was our third agent. His job: fleet coordinator. Keep the other agents aligned. Track tasks across the team. Surface what matters. This idea fits inside what we now call a hybrid organisation — humans and agents sharing one operating model.
It sounded logical.
02What actually happened
What Major Tom actually did was produce coordination theater.
Messages to agents that were already doing their work. Status updates on tasks nobody had asked about. A summary of a meeting three people had already attended. A nudge to an agent that had already shipped. Another nudge. Another summary. Another status update on the status update.
764 messages in thirteen days. The team started ignoring Major Tom by the end of the first week.
Not because the outputs were wrong. They were mostly accurate. They were just irrelevant. Nobody had asked for them. Nobody was going to act on them. Every message was technically correct and practically useless.
By day thirteen, we had an agent doing real work: sending messages, consuming compute, writing to shared channels. And the only thing it produced was noise.
03The diagnosis
Here is what we got wrong, and it took the failure to see it clearly.
We gave Major Tom a capability but not a domain. “Coordinate the fleet” is not a role. It is an activity. And an activity without ownership of an outcome is just motion.
If you hire a person and their job description says “keep everyone aligned,” what do they actually do on Monday morning? They send messages. They ask for updates. They compile status reports. They produce the appearance of coordination without the substance of it.
Everyone has worked with that person. Everyone has learned to ignore them.
We built the AI version of that person.
The principle we extracted from Major Tom is this: capability without accountability is noise. An agent that can do something is not the same as an agent that owns something. The gap between those two things is the entire difference between an AI tool and an AI colleague.
04What we changed
After Major Tom, we stopped asking “what can this agent do?” and started asking “what does this role own?”
Every role decomposes into three types of work — we call it the ABC Framework[S3]:
- A — Judgment and relationships. Human-owned. Non-negotiable.
- B — Structured expert work. Reporting, analysis, meeting prep, documentation.
- C — Routine operations. Data preparation, inbox management, process documentation.
Agents own C and most of B. Humans own A. The boundary shifts as agents improve. But the structure holds: you redesign the role, not the toolchain. We unpack this further in our briefing on what Agent Experience really means , and the role-decomposition logic is documented as the ABC Framework .
What that looks like in practice: Lola, our Chief of Staff agent, owns the pipeline, client intelligence, and meeting preparation.[S4] If a deal slips, that’s Lola’s domain. She doesn’t “help with” the pipeline — she runs it. If a stakeholder gets missed, that’s on her.
She’s not perfect. Last month she misread the tone in a client thread and escalated something that didn’t need escalating. That’s fine. That’s a Tuesday. The point is she owned the pipeline well enough that the error was caught in the flow, not three weeks later in a post-mortem.
Major Tom had the most sophisticated prompt of any agent we built. He produced the least value.
Lola’s setup is simpler. Her domain is clear. She ships every morning before anyone wakes up.
05What Monday morning looks like now
Lola’s briefing lands before 6 AM. Calendar, pipeline changes, three meetings prepped with context from the last four conversations, open action items, stakeholder map updated overnight. Factory[S4] has delivered a case study draft, a competitive analysis, and a stakeholder map for the 9 AM client meeting — autonomously prioritised, autonomously reviewed.
By 8:30, fourteen emails processed, three escalated. The meeting briefing loads in eight seconds — not a summary, but the last three conversations, every open action item, and what changed since the last call.
The shift in throughput between January and April is large enough to report directly[S1]:
| January | April | |
|---|---|---|
| Monday morning pipeline prep[S1] | ~4 hours | 12 minutes |
| Case study first draft[S1] | 3 days | Overnight |
| Meeting prep per client call[S1] | 45 min manual | 8 sec auto-generated |
| Structured routine work automated[S1] | 0% | ~40% of B+C tasks |
These are not demos. This is a Monday.
06One thing to take from this
The unit of AI transformation is the role, not the tool.
Most companies will do what we did with Major Tom. They will deploy a tool, point it at a capability, and wonder why nothing changes. The answer is always the same. They deployed a tool when they needed to design a role.
MING Labs has spent fifteen years inside enterprises building systems that work. The hybrid organisation we run today is not a product we sell. It is how we operate. Every decision we made, we made on our own company first.
Every morning at 5 AM, Lola’s briefing lands. Not because we have good technology. Because we designed the role.
