The volume premium is collapsing across domains: more context, more protection, more stimulation, more authority are all showing negative returns past their optimal dose.

The Pattern

An ETH Zurich research team found something uncomfortable about AI coding agents this week. AGENTS.md files, the context documents developers write to help AI understand their codebase, often make the agents perform worse. More instructions. Worse results.

This is not an AI story. It's a dosage story.

The same week, the U.S. economy lost 92,000 jobs in February, pushing multiple Fed governors toward rate cuts. The protective response is predictable. More liquidity. More stimulus. More intervention. Meanwhile, research on European labor laws shows that employment protection, the policy designed to shelter workers from risk, actually suppresses firm-level risk-taking so thoroughly that the firms stop innovating. Protection becomes the thing that creates the vulnerability it was designed to prevent.

And in the nootropics community, users are documenting what pharmacologists call the Yerkes-Dodson effect at high stimulant doses. Past a certain threshold, the stimulant that sharpened your focus starts degrading your actual output. You feel productive. You are not.

Three domains. Same structure. Every input has an optimal dose, and the penalty for exceeding it is invisible because the system keeps telling you it's working.

If you run a team, this is the most dangerous failure mode you face. Not underinvestment. Overinvestment in the wrong layer. More process, more documentation, more meetings, more tooling. All of it feels like progress. The returns went negative three additions ago.

The Tension

The tension is between felt performance and actual performance. They diverge past the optimal dose, and the divergence is undetectable from inside the system.

The ETH Zurich finding is precise here. Developers who wrote longer, more detailed AGENTS.md files believed they were helping their AI agents. The agents got worse. The context didn't clarify. It confused. The developers couldn't tell because the agents still produced output. The output just degraded in ways that required careful measurement to detect.

This maps exactly onto the stimulant research. At high doses, users report feeling sharper, more focused, more capable. Measured performance tells a different story. The felt-sense and the actual output decouple. You lose the feedback loop that would tell you to stop.

Now scale this to monetary policy. The Fed sees 92,000 jobs lost. The response is more liquidity. But European employment protection research demonstrates that past a threshold, protection suppresses the risk-taking that creates the next generation of jobs. More protection. Fewer new companies. Fewer new jobs. The intervention feels responsible. The economy feels supported. The dynamism dies quietly.

India's refusal to stop buying Russian oil regardless of U.S. pressure is the geopolitical version. American diplomatic volume, more sanctions, more pressure, more statements, has passed its optimal dose with countries large enough to have alternatives. India isn't defying the U.S. out of hostility. They've simply stopped registering the signal. The volume itself became noise.

If you're a founder deciding how much documentation to write for your team, how many approval steps to add to your pipeline, how much context to give your AI tools, this is the trade-off: every addition has a real cost that your own experience of the system will hide from you. The only way to detect the inflection point is to measure outputs, not inputs.

What This Unlocks

The concept I keep coming back to is dose curves applied to organizational design.

In pharmacology, the therapeutic window is the range between "not enough to work" and "enough to cause harm." Below the window, the drug does nothing. Inside, it works. Above, it damages. Every serious medication has a known therapeutic window.

Almost no organizational practice has one.

We add process until something goes wrong, then add more. We add context to AI tools until they hallucinate, then add more context to prevent the hallucination. We add employment protections until firms stop hiring, then add incentives to encourage hiring under the protections.

The Graph-Oriented Generation paper is an example of what happens when someone takes dosage seriously. Instead of feeding AI models more context via vector search, GOG uses deterministic AST traversal. Know the structure. Traverse it. 70% token reduction. Not "less context." The right context, at the right dose, through the right mechanism.

This connects to Tuesday's analysis of constraint design as the new human role. Constraint design IS dose design. The human job is not to provide more input. It's to find the therapeutic window for each input and hold the system inside it.

A separate study found that LLM code quality improves dramatically when users define acceptance criteria first, before generating code. Not more instructions. A tighter frame. The dose of instruction stays low. The precision of the constraint goes up.

If you're building a service business, the practical implication is counterintuitive. Your onboarding documents might be hurting your delivery. Your SOPs might be slowing your team. Not because they're wrong, but because you passed the dose window three revisions ago. The test is simple: remove 30% of the documentation and measure whether output quality changes. If it doesn't change, or improves, you found your answer.

Watching Next

Three things I'm watching to test whether the dose-curve framework holds.

First, whether AI agent benchmarks start measuring performance degradation at high context volumes. The ETH Zurich paper opens this door. If other labs replicate the finding, we'll see a shift from "how much can the model handle" to "what's the optimal context dose for this task." Watch for the word "therapeutic" appearing in ML papers by Q3.

Second, whether the Fed's rate cut response to job losses accelerates or slows new business formation. If European labor rigidity research is predictive, more monetary protection should correlate with fewer new firms, not more. The BLS startup data for Q2-Q3 will tell the story.

Third, in your own work: pick one process you've been adding to. Stop adding. Measure output for two weeks. If output holds or improves, you've found a dose window. If it degrades, you were still in the therapeutic range.

Underweighting

I might be wrong about the universality of the dose curve. There's a strong argument that some inputs have much wider therapeutic windows than others. Capital, for example, might degrade at a much higher dose than context or process. The AI valuation narrative shift from "bubble" to "legacy software goes to zero" suggests that capital deployed toward AI may have a wider window than I'm implying.

I also might be imposing the cross-domain pattern rather than discovering it. The ETH Zurich finding is about a specific kind of context (AGENTS.md files) in a specific system (AI coding agents). The jump from there to monetary policy to stimulants is structural reasoning, not empirical proof. The pattern feels real. Felt-sense and measured reality diverging is literally the pattern I'm warning about.

Octavia Butler wrote that God is Change. Not that God changes things. That Change itself is the fundamental force. Maybe the dose curve isn't the frame. Maybe the frame is that every system resists the recognition that what worked at one scale stops working at another. We keep dosing because change feels like failure.

Bottom Line

Every input to every system has a therapeutic window. Below it, nothing happens. Inside it, the system performs. Above it, the system degrades while still reporting that it's working. The penalty for overdose is invisible to the person administering it. If you build anything, your job is not to add more. It's to find the window, then defend it.