The fear, and what is actually happening
Most people now expect AI to take jobs — 71% of Americans, in a Pew survey published this month, up seven points in two years. The fear is reasonable. The prediction is mostly wrong, and wrong in an instructive way.
AI doesn’t replace jobs. It takes tasks. Goldman Sachs’ estimate is that current AI could automate work worth about a quarter of US working hours; the IMF puts about 60% of jobs in advanced economies as “exposed”, roughly half of them likely to benefit. Look inside any role and the pattern is the same: the digital parts — the drafts, the research, the follow-ups, the reconciliations, the summaries — are increasingly things a machine can do. The rest of the role is not.
So the honest sentence is not “AI will replace most people”. It is: AI will take over parts of most people’s digital workload. Which raises the question nobody on either side of the fear is answering: then what happens to the rest?
Why real goals stay yours
When Harvard and BCG put AI in the hands of 758 consultants, the ones using it finished 12% more tasks, 25% faster, at 40% higher quality — and the lower performers gained the most. Humans equipped with AI beat both humans alone and AI alone. But the same study found a jagged frontier: on tasks just outside the machine’s competence, the AI-equipped consultants did worse than the unaided ones, because they trusted output they couldn’t judge.
METR, which measures how long a task an AI agent can finish on its own, finds the horizon doubling roughly every seven months — and finds the results bimodal. On tasks that take an expert 90 minutes to three hours, GPT-5 “succeeds 100% of the time for around one-third of the tasks, fails 100% of the time for around one-third.” Not “usually works”. Brilliant here, hopeless there, with no visible seam between the two.
Ethan Mollick, who wrote the book on working with AI, published a piece in June titled Co-Existence and the End of Co-Intelligence. His conclusion: working with something “sometimes better than you, and sometimes hilariously worse, is not a problem you solve once. It is a relationship you negotiate.” His open questions are the human ones — when should you refuse AI’s help, even when it is offering? When should you hand over the keys entirely?
That is the shape of it. Real goals are messy. They involve context the machine was never told, judgment about which of two bad options is worse, relationships with people who remember what you said last month, trade-offs that only make sense from inside your life, and consequences you — not the model — will live with. Those don’t stay with you because the AI is weak. They stay with you because they are yours.
The question is not whether AI replaces you. It is how you and the AI divide the work — and who decides.
What the builders learned this year
In February, Mitchell Hashimoto described a habit: every time his coding agent made a mistake, he engineered a permanent fix into the agent’s environment rather than into the prompt. He called it engineering the harness. Within weeks OpenAI and Anthropic had published their own accounts, and the formula the field settled on is now everywhere:
Agent = Model + Harness.
The model is the raw intelligence. The harness is everything that turns it into something that actually gets work done: which tools it may call and when, memory that persists past one conversation, a sandbox where mistakes are contained, context management so it isn’t drowning, and guardrails — scoped permissions, approval steps, verification loops, observability. The standard definition puts it plainly: the harness “handles what the model cannot: long-running state, sandboxed execution, feedback loops, and recovery from failure.”
Anthropic’s own usage data shows what a harness does to the human’s job. In its June Economic Index, the median blog post written in chat took thirteen rounds of back-and-forth. The median blog post written in Claude Code — the same models, wrapped in a harness — took one prompt. The model did not get smarter between those two columns. The harness decided how much a person had to steer.
Developers now have this. Everyone else still has a chat box.
The missing harness is the one around you
A blank prompt is an expert tool sold to non-experts. It asks you to already know what to ask for, in the right words, and to check what comes back yourself. That is why AI today mostly pays off for the people who already knew how to drive it — and why someone with a real ambition and no technical background so often has nothing to show for it. Not knowing where to start is not a skill gap. It is a missing harness.
A harness for a person is not a harness for a codebase, but it has the same parts:
- The goal is the unit, not the task. You say what you want. It works out what the tasks are — a dated plan you can see, with milestones — instead of waiting for you to know what to ask.
- Tool dispatch becomes background work. Every task is classified: digital, action, or decision. The digital ones run automatically in the background while you’re doing other things.
- Memory holds only what you’ve shared. Your goals, journal and decisions are the context it works from — scoped to you, exportable, deletable.
- Approval steps become confirm gates. Nothing acts in your name without your agreement.
- Verification loops become checked work. Nothing reaches you as raw model text; outputs pass schema checks and a verifier before they land, and background work arrives for you to accept or send back.
- Recovery becomes the hold. When a task genuinely needs something only you know, it stops and asks — once — rather than guessing.
- Termination becomes what comes back to you: the decisions, and the real-world steps.
That is the whole design of iSpirits Cloud, and it is one sentence long: it works out the tasks, runs the digital ones in the background, and brings you only the decisions and the real-world steps. A few minutes of chat or voice a day is what it costs you.
Who holds the reins
The word “harness” invites a bad reading, so let us be exact about it. A harness is worn by the thing doing the pulling. The reins are held by the one who decides where to go.
The AI wears the harness. You hold the reins. You keep exactly three things — the goal, the decisions, and the real-world actions — and even the project management is delegated. Not an AI that replaces you. Not you as its assistant, either. A right hand you raise, that runs the digital work and knows only what you’ve chosen to share.
You bring the ambition. It runs the rest.
What this does not fix
- It cannot want the goal for you. A harness makes a stated ambition executable; it does not supply one.
- The frontier is jagged and it will be wrong sometimes — that is precisely why every output is checked and reviewable, and why nothing acts in your name without agreement. Treat what it brings you as a draft from a capable colleague, not a verdict.
- Real-world actions stay real-world. It will prepare the call; it cannot make it.
- Background execution is rate-capped on the free plan (two runs per 24 hours; five on paid plans). We are pre-launch: there are no customer numbers in this piece, and the only product figures are ours.
Your iSpirit is the right hand you raise: it runs your goals and daily tasks, learns only what you choose to share, and grows with you — so every month it works better than the last. A few minutes of chat or voice a day.