Six AI tools. One human connecting all of them.
A product manager at a mid-size startup described their current weekly workflow: Claude for brainstorming, ChatGPT for rewriting product specs, Cursor for implementation work, Perplexity for research, Notion AI for documentation, and a tool called Atoms AI for larger tasks. None of these replaced their work. They just moved where the work happens.
"None of these tools actually replaced my work," they wrote. "They just redistributed it. I'm still the one dragging context between all of them."
This is probably the most honest description of AI-augmented knowledge work I've seen this year.
The context-dragging problem
Each of these tools is stateless relative to the others. Claude doesn't know what ChatGPT rewrote. Cursor can't see the Perplexity research. Every handoff between tools requires a human to summarize, copy-paste, reframe, and re-prompt. The AI handles execution within its own context window (the amount of text it can hold in memory at once). The human handles everything that crosses tool boundaries.
The result is that the human is now functioning as an integration layer - which is a real job, but not necessarily one that appears on any job description or generates any credit during a performance review.
This scales badly. A solo developer using Claude Code in a well-documented codebase has genuinely fewer interruptions than they did two years ago. A solo content creator using one AI writing tool produces more per hour. The gains are real in bounded domains.
Cross two tools, and you're managing interfaces. At six tools, you've built yourself a part-time job as a systems integrator.
Why the tools can't solve this yet
Most AI assistants don't share memory or state with each other. There's no standard protocol for a Claude conversation to hand off to a Cursor session with full context intact. These tools were built by competing companies with no incentive to interoperate cleanly.
Some AI agent frameworks are trying to fix this - systems where one AI orchestrates others, automatically passing context across tool boundaries. But in practice, most of these are still fragile and require significant setup. The product manager juggling six tools manually is using what actually works today, not what's theoretically possible.
Some of the problem is also structural. Complex work involves multiple domains - research, writing, code, documentation - and moving between domains has always required a human judgment call about what context matters and what can be dropped. That judgment call isn't something current AI tools automate well. They're good at working within a domain. The between-domain coordination remains yours.
AI tools are making each domain faster. They haven't closed the gap between them. Until they do - through better integrations, shared memory, or some kind of agent-to-agent handoff standard - the person in the middle stays very busy.