yarnnnyarnnn
← Back to blog

The Messy Middle of AI Work: Why the Future Isn't Tools or Full Autonomy

·9 min read·Kevin Kim

The AI conversation right now is polarized in a way that's actively unhelpful. On one side: AI is a tool, use it like you use Excel, nothing fundamentally changes. On the other side: AI replaces everything, white-collar work collapses by 2028, start hoarding canned goods. The tool crowd refuses to acknowledge the cracks forming under their feet. The doomer crowd assumes the cracks mean the whole floor is about to give way.

I think both camps are wrong — not because the truth is somewhere in the middle (that's lazy), but because they're both skipping an entire era of work that we'll actually live in for the next decade.

Three Eras, One Transition Nobody's Building For

Here's how I think about the trajectory:

Era 1: Human-Driven Work. AI as autocomplete. You prompt ChatGPT, it responds, you copy-paste the useful parts into your actual work. The human does the thinking, the structuring, the deciding. AI saves time on first drafts. This is where most knowledge workers still are in early 2026.

Era 2: Human + AI (Supervised). AI handles execution — drafting, research, synthesis, monitoring — while humans handle judgment. What to produce, who it's for, whether the output meets the bar. The AI gets better at execution over time because it accumulates context about your specific work. The human stays in the loop not because the AI can't act, but because the decision layer still requires human accountability. This is the messy middle.

Era 3: Agent-to-Agent. Fully autonomous agents negotiating with each other. Your AI talks to your client's AI. Agents coordinate workflows, execute transactions, resolve dependencies — humans as stakeholders, not operators. This is where the hype points.

The discourse is stuck debating Era 1 versus Era 3. The incumbents say "AI is just a tool" because acknowledging Era 2 means admitting their products need fundamental rearchitecting. The accelerationists skip straight to Era 3 because the messy middle isn't exciting enough for a viral Substack post. Meanwhile, almost nobody is seriously building for Era 2 — the era we're actually entering.

The Cracks Are Real. The Timeline Is Not.

I'm not going to pretend the transition isn't happening. Claude Code is already reshaping how SaaS companies get valued — the stock market volatility around AI-driven displacement isn't hysteria, it's repricing. The cracks in the existing knowledge work economy are visible. Junior analyst roles that took 40 hours now take 4. Code review cycles are compressing. Research that required a team of three gets done by one person with the right agent setup.

But here's what the doomer framing misses: cracks don't tell you the shape of the break, and they definitely don't tell you the timing.

Existing data moats don't evaporate overnight. Distribution advantages don't disappear because a new tool exists. Switching costs are real — the enterprise that runs on Salesforce isn't going to rebuild around an AI agent next quarter, no matter how good the demos look. Regulatory posture varies wildly by geography. Social norms around AI trust are still forming. The EU, the US, Japan, Korea — each jurisdiction is adapting to AI at a different pace, with different constraints, and those differences create friction that slows any uniform "revolution."

The cracks in the sand are there. But sand shifts slowly and unevenly.

What Adoption Curve Are We Even On?

This is the question I keep coming back to, and I don't think anyone has a good answer.

Are we on the electricity curve? The grid existed for 40 years before factories were actually redesigned around it. The first generation of factory owners just replaced steam engines with electric motors in the same layout — same floor plan, same workflow, just a different power source. The real productivity revolution came a generation later, when new factories were purpose-built for distributed electric power. If AI follows this pattern, we're in the "electric motor bolted onto a steam factory" phase. The real transformation is a decade away, and it looks nothing like what we're building today.

Are we on the internet curve? Massive hype, spectacular crash, then the actual transformation happened quietly over the following decade. The companies that won — Google, Amazon — weren't the ones dominating the conversation in 1999. If AI follows this pattern, the current crop of AI startups (including mine) might be the Pets.com generation, and the real winners haven't been founded yet.

Are we on the smartphone curve? Faster adoption, but the app economy that actually mattered took five years after the iPhone launched. The device was ready before the ecosystem was. If AI follows this, the models are the iPhone — already here, already good — but the ecosystem of AI-native workflows is years from maturity.

Or are we on the containerization curve? Malcom McLean's shipping container was technically simple. But ports, labor unions, trade regulations, international standards, insurance frameworks — the entire logistics system took 20 years to reorganize around a metal box. The technology was ready long before the world was. If AI follows this pattern, the constraint isn't model capability. It's everything else.

I genuinely don't know which curve we're on. I suspect it's a blend — fast in some sectors (software, content, analysis), slow in others (legal, healthcare, government), and unpredictable where regulatory and social friction create bottlenecks nobody modeled. Trying to predict the timing is dangerous. But not thinking about it at all is worse — because the architecture you build today determines whether you survive whichever curve we're actually on.

Agent-to-Agent: The Era Everyone Assumes and Nobody's Building the Rails For

Here's where the conversation gets genuinely uncomfortable.

Agent-to-Agent work — Era 3 — is the logical endpoint of everything the AI industry is building toward. Google's A2A protocol. MCP tool use. Multi-agent orchestration frameworks. The vision is clear: your AI agent negotiates with your client's AI agent, coordinates with your team's AI agents, executes workflows across organizational boundaries without human intervention.

It's also the era that requires an entirely new economic infrastructure that doesn't exist yet.

Think about what Agent-to-Agent work actually requires. Not the model capability — that's arguably close enough. The everything else:

An agent economy. How do agents transact with each other? Who pays when Agent A requests a service from Agent B? What's the settlement layer? Today's payment rails — credit cards, ACH, wire transfers — are built for humans and human-controlled institutions. An agent executing thousands of micro-transactions per day needs something fundamentally different.

Trust and verification. When two agents negotiate an outcome, who audits the decision chain? How do you verify that Agent A's recommendation to Agent B wasn't based on hallucinated data? The trust frameworks we have — contracts, compliance reviews, audit trails — assume a human is accountable at every decision point. Remove the human, and the entire trust architecture needs rebuilding.

Governance. Who's liable when an agent-to-agent workflow produces a bad outcome? If my AI agent tells your AI agent to execute a financial transaction that turns out to be wrong, which party bears responsibility? Current legal frameworks have no answer. They weren't designed for autonomous actors negotiating on behalf of principals who may not have reviewed the specific decision.

Regulatory rails. Finance, accounting, healthcare, legal — these sectors run on compliance frameworks built for human actors. GAAP doesn't have a standard for "AI agent produced this financial statement." HIPAA doesn't contemplate an agent accessing patient data to coordinate with another agent at a different institution. SOC 2 audits don't cover agent-to-agent data flows.

These aren't edge cases. They're the foundation. And none of it is built.

The Revolution We're Not Talking About

The more I think about Era 3, the more I believe the actual work economy revolution isn't about making agents smarter. The models are smart enough. The actual revolution is building the economic, legal, and trust infrastructure that allows autonomous agents to operate in the real economy.

That's not a model problem. It's not an engineering problem. It's a civilization-scale coordination problem — the kind that historically takes decades, not quarters.

Which brings me back to the messy middle. Era 2 isn't a waypoint you rush through on the way to Era 3. It might be the era where the most important work happens — not the AI work itself, but the building of trust, governance, and economic rails that make Era 3 possible eventually.

The question I can't stop asking: are we so focused on making agents that can do the work that we've completely ignored building the world those agents would actually operate in? And if so — is that infrastructure, not intelligence, the real bottleneck?

I don't have the answer. But I think it's the question that matters most right now, and almost nobody in the AI conversation is asking it.

References