The five stages of AI, and where the money actually is
Hey, it's Cindy 🌱 You commented STAGES, so here is the full map: what each of the five stages actually means.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
the map
The five stages, defined properly
The stages are not "smarter model" steps. Each one is a change in who decides what happens next, and that is the only definition that stays useful as models improve.
💬 Stage one, large language models. You ask, it answers, one turn at a time. Claude, ChatGPT and the rest of the chat assistants. You are still the hands: every file that gets moved, every tool that gets opened, every step of the actual work happens in your tabs. This is where the overwhelming majority of people have stayed since 2023.
👩💻 Stage two, agentic AI. You give it a job instead of a question. Anthropic's engineering team draws the line precisely here: agents are "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks" (Building effective agents). Claude Code is the everyday example: it reads real files, runs real commands, and hands you a finished result to review.
🌱 Stage three, multi-agent systems. One lead agent splits the work and delegates to several sub-agents running in parallel, then merges what comes back. One researches, one drafts, one checks the work. This is not theoretical, it is a documented Claude Code feature (sub-agents docs), and it is how I run my own scripts every week.
💡 Stage four, AGI. AI that matches human cognition across essentially everything, rather than being strong in narrow bands. Nobody agrees on the definition, the test, or whether current approaches even lead there. Treat any confident claim about stage four, in either direction, as an opinion wearing a lab coat.
🔎 Stage five, superintelligence. Systems that exceed human reasoning outright. This one is entirely hypothetical, which has not stopped it becoming the loudest part of the conversation, and it is worth knowing why that is (below).
💡 The one-line version. Stage one explains. Stage two executes. Stage three delegates. Stages four and five are forecasts, not products. Moving one real task from stage one to stage two changes your week more than any model upgrade will.
be honest
Where we actually are
Stages one, two and three all exist and you can use every one of them today. Stages four and five are contested, and the actual expert numbers are much less dramatic than the headlines.
Researchers are broadly sceptical In the AAAI 2025 Presidential Panel survey, 76% of respondents said scaling up current approaches is unlikely or very unlikely to produce AGI. That survey ran across 475 participants, mostly academic.
But timelines did move fast A survey of 2,778 published AI researchers put a 50% chance of machines outperforming humans at every task by 2047, and 10% by 2027. The 2047 figure was 13 years earlier than the same survey a year before.
Stage five is a policy fight The Future of Life Institute's statement calling for a prohibition on superintelligence development until there is scientific consensus and public buy-in has gathered tens of thousands of signatures. That is a governance debate, not a product roadmap.
What that adds up to Serious people disagree by decades on stage four and cannot agree on a definition for stage five. Meanwhile stages two and three are sitting in a $20 subscription, mostly unused.
Why this matters for your decisions. Content that leads with stage four and five is optimised for how it feels, not for what you can do on Monday. If a forecast would not change a single action you take this month, it is entertainment. Read it as such and then go build something at stage two.
the actual answer
Where the money actually is
Not in predicting stage four. The money is in the gap between what stage two and three can already do and how few people can operate them, and there is real data showing that gap widening.
Anthropic's Economic Index tracked how people actually use Claude and found directive usage, meaning full task delegation rather than back-and-forth collaboration, rose from 27% of Claude.ai conversations in December 2024 to 39% by August 2025, the first period where automation overtook augmentation. Among API customers, where businesses build the things they sell, 77% showed automation patterns (Economic Index report). The demand is moving towards delegation, and delegation is stage two and three work.
The second thing worth knowing is that stage three is not automatically better, and this is where the paid skill actually sits. Anthropic's own multi-agent research system beat a single agent by 90.2% on breadth-first research queries, and it did so while burning roughly 15 times the tokens of a normal chat (how we built our multi-agent research system). They are blunt about the limits too: domains "that require all agents to share the same context or involve many dependencies between agents are not a good fit."
🧠 The skill nobody is selling yet. Anyone can spin up five agents. The valuable judgement is knowing which tasks justify the cost and which are better done by one agent with good instructions. That single decision is the difference between a system that saves money and one that quietly burns it, and it is exactly the thing clients cannot evaluate for themselves.
stage one to two
Stop asking, start delegating
The jump is not technical. It is a change in what you type. A stage-one request asks for information. A stage-two request describes a finished outcome and lets the model plan the steps.
🔎 Pick a task you have done three times this month. Repetition is the qualifier. Something you have never done manually is a bad first delegation, because you will not be able to tell whether the output is right.
📋 Give it the real inputs. Attach the actual files, or connect the tool that holds them under Customize, then Connectors. An agent with no access to your real material is still a chatbot with extra steps.
🌱 Ask for the artefact, not the instructions. "Tell me how to do X" is stage one. "Here are the inputs, give me the finished X" is stage two. That is genuinely most of the shift.
💡 Review the work, not the summary. Read what it actually changed rather than the confident paragraph it writes at the end. This one habit separates people who trust agents appropriately from people who get burned once and quit.
👩💻 your first real stage-two task
Here are the raw inputs for [the thing you do every week, for example "our Monday performance summary"]: [attach the files or point at the folder]. I want the finished piece of work back, not a set of instructions for making it myself.
Work through it end to end: read every input properly, pull out what actually matters, and produce the final artefact in the same structure as the previous version I have attached. Where something is missing, ambiguous or does not reconcile, flag it explicitly instead of filling the gap with a plausible guess. When you are done, list every assumption you made and tell me which parts you are least confident about, so I know exactly where to look first when I review it.
stage two to three
Split the work, then make it check itself
Stage three is worth reaching for when a job has genuinely independent parts that can run at the same time, and when the output is big enough that one context window struggles. Research, competitor sweeps and content pipelines fit. Most coding and anything where every step depends on the last does not.
🔎 Test that the parts are actually independent. If sub-task B cannot start until it sees the result of sub-task A, you do not have a parallel job, you have a sequence. Run it as one agent and save yourself the token bill.
📋 Give each agent one narrow brief. Vague sub-agents produce overlapping, duplicated work. "Research X across these three sources and return quotes with links" is a brief. "Help with research" is not.
💡 Add a checker that can actually fail. The step people skip. A verification agent that can only say "looks good" is decoration. Give it a test that has a real pass or fail condition: a count that must match, a link that must resolve, a claim that must appear in the source.
🧠 Read the first few runs properly. A finished run means the session completed, not that the work was right. Open two or three end to end before you let the system run unattended.
🌱 run a stage-three job with a real check
I want you to run this as a multi-agent job rather than doing it yourself in sequence: [describe the task, for example "research how five named competitors position their pricing"]. First tell me whether this task genuinely benefits from parallel agents or whether one agent would do it better, and say so honestly if the answer is one agent.
If it does benefit, split it into independent sub-tasks with no dependencies between them, give me the specific brief you would hand each sub-agent, and run them. Then take the merged output and verify it before you show me anything: define in writing the specific check that proves the result is right, including what a failure would look like, run that check, and paste the raw result. If the check fails, change your approach and run it again rather than explaining the failure away. Finish with the evidence and a note on what the parallel approach actually cost compared to doing it in one pass.
The cost gate, said plainly. Stage three burns roughly 15 times the tokens of a normal chat, so on a capped plan a few careless multi-agent runs will eat your allowance in an afternoon. Use it for jobs where the answer is worth that, and use a single well-briefed agent for everything else.
where are you
Which stage are you actually on?
Not a vibe check. Look at what happened after your last three AI sessions.
Stage one You got text back and then did the work yourself. Nothing outside the chat window changed.
Stage one and a half You have connectors switched on and a paid plan, but you still ask one question at a time. The capability is there, the habit is not.
Stage two You described an outcome, it touched real files or tools, and your job was reviewing rather than doing.
Stage three Several agents worked in parallel on parts of one job, and something verified the result against a check that could have failed.
🌱 The move this week. Take one task you did three times this month and push it exactly one stage up. One task, not all of them. Nobody has ever regretted moving too carefully here, and plenty of people have built stage-three systems on top of work they never validated at stage two.