Garry Tan recently described the "harness as the product". The magic that makes a great agentic experience isn't just the language model, it's the wrapper that gives it just the right context at just the right time and minimizes noise. The bottleneck isn't raw intelligence, it's architecture.
The same principles govern human cognition, a parallel directly relevant to engineering leaders managing the transition to agentic development.
Working memory is the harness
Tan argues for thin harness, fat skills. As it turns out, human intelligence runs on a thin harness too; limited by the seven items (George Miller, 1956) or four chunks (Nelson Cowan) that we can hold in working memory.
What makes a 100x developer different from a novice is not a bigger working memory, it's richer chunks. A chess grandmaster doesn't see thirty-two pieces; they see five or six meaningful configurations. A senior engineer doesn't see lines of code; they see architectural patterns. Each chunk encodes enormous compressed knowledge, loaded from long-term memory in a single retrieval operation. The harness stays thin. What flows through it becomes fat.
The anti-pattern in AI systems (forty tool definitions eating half the context window) is precisely what happens when a novice tries to function without proper chunking. They put everything in working memory because they haven't yet built the long-term structures that would let them retrieve the right knowledge at the right moment. Fat harness, thin skills. It's inefficient in software. It's exhausting in humans.
Stress: when the harness gets hijacked
Human cognition differs from AI in an important way, with drastic negative impact on most engineering teams.
Daniel Kahneman and Keith Stanovich described two modes of thinking: System 1 (fast, associative, pattern-matching, prone to certain errors) and System 2 (slow, deliberate, rule-following, reliable for logical operations). Under normal conditions, problems are routed to the appropriate system. Under stress, the routing breaks down.
High stress consolidates attention in the limbic system, the part of the brain that governs survival responses. This is an effective strategy for avoiding a predator, but counterproductive when debugging complex problems or devising creative solutions. The prefrontal cortex, which handles those tasks, goes partially offline.
This isn't only a human problem. A recent study in Nature's digital medicine journal found that emotion-inducing prompts can elevate measurable "anxiety" responses in large language models, and that mindfulness-style prompting reduced those responses. The parallel is imperfect but striking: context shapes performance in both biological and artificial systems.
Context windows and rumination
There's a second, subtler way that stress degrades human cognitive performance, one with a direct AI analogue.
When the brain's Default Mode Network malfunctions, it produces rumination: a mental state in which irrelevant thoughts, unresolved concerns, and competing priorities flood working memory and crowd out the task at hand. The engineer who sits down to solve a technical problem but keeps cycling back to an unresolved conflict with a stakeholder, a looming deadline, or an ambiguous piece of feedback is experiencing something functionally similar to a context window stuffed with the wrong information.
A bloated CLAUDE.md file doesn't make the model smarter. It makes the model's attention degrade as it attempts to weight too many competing signals simultaneously. A leader carrying too much unprocessed cognitive overhead into a deep work session faces the same problem. The right information at the right moment is as important for human performance as it is for AI performance.
Resolvers and retrieval cues
The problem of expertise is not knowing things. It's getting the right knowledge into working memory at the right time. A doctor who tries to remember every symptom in isolation is like an LLM consulting a RAG system for context. Semantic search is looking for ideas, but at the wrong granularity, and the records retrieved provide disparate facts, not elegant schemas that illuminate a particular problem.
Instead, experts structure their knowledge to optimize both context and retrieval cues in a vastly more effective way than RAG systems operate today. An engineering leader will immediately pattern-match a team sandbagging because they don't trust another part of the organization to do their part, or an engineer writing non-performant code because they are taking product management direction too literally. They aren't finding a fact; they're seeing a pattern.
Novices take notes; experts build mental models. The difference is integration: the ability to connect new information to existing schemas to create emergent understanding. RAG is like high-volume note-taking. It can find relevant passages, but doesn't reason across them to create new knowledge.
Deliberate practice as the learning loop
Sitting at a piano hitting random keys for ten thousand hours does not produce a musician. Deliberate practice does. What separates elite performers from competent ones is not raw hours, but the quality of the feedback loop: structured engagement, specific targeting of the gap between current and target performance, and immediate corrective feedback that rewrites the schema.
The learning loop in software (such as Andrej Karpathy's autoresearch) performs a similar function. Clearly define an evaluation function, try many meaningful experiments, and learn from the best results.
Whether human or machine, the learning loop is what optimizes performance. We can learn to structure our experiences to optimize our own learning, and thoughtfully create such learning loops to optimize our agents.
What this means for humans
Just like agentic systems, we can optimize our cognition and productive output. Here's how:
- Understand your own neural architecture. Performance is shaped by physiological and psychological factors that most engineers never examine: sleep quality, stress load, context-switching costs. These aren't soft concerns; they are system optimization parameters.
- Optimize for retrieval, not just storage. Memorizing more facts is less valuable than developing elegant mental models that explain real-world problems and solutions. Build schemas, not inventories; then test and restructure.
- Route problems to the right system. Most of the challenges an engineer or leader faces require System 2: deliberate, prefrontal, logical processing. Stress and cognitive overload degrade that capability. When you notice the prefrontal cortex start to go offline, take a deep breath and consider what you need to fully recruit its capability.
- Manage attention like a scarce resource. Fragmented attention is how engineers and leaders lose sight of the forest. You are still in meetings, still responding to Slack, and increasingly performatively reviewing agentic outputs, but you've lost perspective on what's important. Deep work is necessary to be productive, but more importantly, to stay oriented in a fast-changing world.
- Build your own learning loops. Get deliberate in retros. At the end of a sprint or a project, reflect on not just what went well and what didn't. Focus on what was just "OK", and how it could have been excellent. In autoresearch terms: be really thoughtful about selecting evaluation functions, then deploy agentic learning loops wherever you can.
The architecture Tan argues for emerged from the same constraints that shaped biological intelligence. Implementing these models in silicon both increases the demands on human intelligence and is an exercise in optimizing both.
References
- Miller, G. A. (1956). The magical number seven, plus or minus two. Psychological Review, 63(2), 81-97. doi:10.1037/h0043158
- Cowan, N. (2001). The magical number 4 in short-term memory. Behavioral and Brain Sciences, 24(1), 87-114. doi:10.1017/s0140525x01003922
- Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux. openlibrary.org
- Arnsten, A. F. T. (2009). Stress signaling pathways that impair prefrontal cortex structure and function. Nature Reviews Neuroscience, 10(6), 410-422. doi:10.1038/nrn2648
Originally published as an article on X, April 2026. Related: The Real Reason 71% of CEOs Feel Like Frauds, these mechanics under social threat.