4 min read
Two of the biggest AI labs picked the same fortnight to declare that chatting with a bot is over, and that “work” is the new interface. Anthropic pushed Claude Cowork out of the desktop app and onto phones and the web, and OpenAI answered days later with ChatGPT Work. The workflow both are selling, and that has flooded every feed this week, is simple to describe and seductive to try: instead of asking the AI a question, you hand it an entire goal and walk away.
The workflow, in plain terms
Here is the loop everyone is posting. You give the agent a project, not a prompt: “pull the numbers from these three files, build a board deck, and draft the email to send it round.” The agent connects to your apps and files, breaks the job into steps, grinds away for minutes or hours, and hands back a finished artefact, a populated spreadsheet, a slide deck, a shareable web page, rather than a wall of chat text you then have to copy-paste into shape. On the developer side, the same idea shows up as Claude Code subagents: spinning up several agents in parallel, each in its own isolated context, to chew through a codebase-wide migration or a big security audit all at once.
Why it is genuinely good
When it works, it collapses the annoying gap between “the AI told me what to do” and “the thing is actually done.” For multi-step drudgery with a clear finish line, gathering, formatting, cross-referencing, drafting, it is a real step up. The parallel-subagent trick in particular is legitimately powerful for large, splittable jobs where five workers genuinely beat one.
Now let me poke some holes
The demos are glorious. The reality has sharp edges. Three worth knowing before you go all in:
It burns tokens like a bonfire. An agent that “works for hours” is an agent generating and re-reading enormous amounts of text, and you pay for every one of those tokens. Spin up five subagents and you have five times the context and five times the bill. A task that costs pennies as a single sharp prompt can cost real money as a swarm.
It fails in expensive ways. A chatbot that gets it wrong wastes a reply. An agent that misunderstands the goal can spend an hour and a fortune confidently building the wrong deck, and you only find out at the very end. The longer the leash, the bigger the mess when it wanders.
Most tasks do not need it. The viral clips are cherry-picked hero runs. A huge share of what people are throwing at these agents is a five-minute job dressed up as an autonomous mission.
The leaner alternatives nobody is posting
Before you delegate your whole life to a swarm, try the cheap version first.
One tight, well-scoped prompt. Nine times out of ten, a single clear instruction with the right context attached gets you 90% of the result for 5% of the tokens. Agents are for when you genuinely cannot specify the steps up front, not for when you cannot be bothered to.
A smaller, cheaper model. Reformatting a spreadsheet or drafting a routine email does not need a frontier model straining for an hour. Point the boring work at a cheap, fast model and save the expensive brain for the genuinely hard step.
A saved template or a two-step chain, not a five-agent swarm. Most “agentic workflows” people show off are really just a prompt, a document and a bit of patience. If you can write the steps down yourself, you probably do not need an autonomous agent to rediscover them at a couple of dollars a run.
None of this means the “give it the whole project” workflow is hype. It is real, and for the right job it is superb. But the version going viral this week is the maximalist one, because maximalism makes a better video. The smarter move, as usual, is to reach for the smallest tool that actually finishes the job.
The AI tool stack actually worth paying for
One email a week. The tools, models and moves that matter, minus the hype and the horseshit filter set to maximum. Free.
Did you know: tools like n8n and simple scripts have been doing “agentic workflows” for years, deterministically and for a fraction of the cost. A lot of what an AI agent does with expensive reasoning, a boring if-this-then-that automation can do for free, and it never once hallucinates the wrong spreadsheet. See n8n’s own take on how the tooling is shifting.