AI Has Changed the Meaning of Productivity

 

Your engineering team shipped 40% fewer lines of code this quarter.

Your CFO sees the numbers, panics, and asks: "What's happening?"

Everyone has different answers.

The engineer says the code is better. The sales team says their deals are bigger. Operations says they handled twice the volume with the same headcount. The CFO wants a number that makes sense.

None of them are wrong. All of them are measuring productivity the old way.

---

The Problem: We're Measuring the Wrong Thing

For over a century, productivity meant the same thing: output per unit of input.

Lines of code per hour. Sales calls per day. Reports generated per employee. Reports per analyst. The formula was simple: more input meant more output.

AI broke this equation in a specific way. It made generation cheap and verification expensive.

A developer using GitHub Copilot can generate a thousand lines of code per hour. But someone still has to read every line, understand the logic, catch the bugs, and take responsibility for the outcome. The work moved upstream.

A marketer using AI can draft fifty email variations in minutes. But each variation requires testing, judgment about brand voice, and decisions about which one matters. The manual work didn't disappear; it changed shape.

A lawyer using AI research tools can review ten times more case law. But someone still has to read it critically, apply judgment, and decide which precedent applies to this specific case. The friction didn't decrease; it shifted.

The old metric (output per hour) sees this as productivity decline. The CFO's dashboard goes red. The manager looks busy. But the actual value moved in a different direction entirely.

---

The Shift: From Creator to Curator

The most productive person in your organization used to be the one who could execute fastest.

Write the most elegant code. Close the most deals. Produce the most reports. The bottleneck was human execution, so you optimized for execution speed.

AI removes that bottleneck.

Now the most productive person is the one who can evaluate AI output with the sharpest judgment. They ask better questions. They spot flawed assumptions in AI-generated analyses. They know when machine output is exactly right and when it's confidently wrong.

This is not a small change. It inverts what capability matters.

A coder who can write 500 perfect lines from scratch is valuable. A coder who can generate 5,000 lines with AI and then evaluate which ones actually matter is 10x more valuable.

A strategist who can synthesize research into a memo is valuable. A strategist who can prompt AI to generate ten memo variations, then judge which framing will actually move the organization is 5x more valuable.

Productivity shifted from execution speed to judgment quality.

 

---

The Hidden Cost: The Verification Tax

Here is what the productivity dashboards do not capture.

MIT studied this. A writing task that took 10 hours now takes 6 hours with ChatGPT. The worker feels 40% more productive. But someone still has to read the output. Fact-check the numbers. Verify the analysis. Evaluate the tone. That verification work does not disappear. It just moves to the reviewer, or stays with the writer in the form of obsessive quality-checking.

A 2026 study from BetterUp Labs found that 41% of workers had received "workslop" (content that looks productive but requires rework). Each instance cost roughly two hours of cleanup work. The annual cost to a 10,000-person company: $9 million.

The productivity gain (generation time saved) was real. The hidden cost (verification time required) was not counted.

Your company is shipping work 40% faster. But someone is spending 30% of their day verifying it is not hallucinated. The net gain is not 40%. It is closer to 10%. And nobody is measuring the 30%.

---

Where This Goes Wrong: The Burnout Paradox

AI makes work faster. It does not make it lighter.

A landmark case study from UC Berkeley and Yale followed a 200-person tech company after introducing generative AI. Adoption was voluntary. Employees loved it. They felt hyper-capable, able to tackle more tasks at higher speed.

But the recovered time did not stay recovered. Because AI made starting tasks effortless, workers filled every recovered minute with new work. The natural micro-breaks (the walk to get water, the moment of staring out the window between tasks) disappeared. Workers extended their workloads without being asked. Only 8% of the time saved by AI was reinvested in their own recovery.

The result: 45% of frequent AI users now report burnout, versus 35% of non-users. The speed increased. The expectations increased faster. The capacity gap widened.

The CFO's dashboard showed productivity up 40%. The employees' experience was chaos compressed.

---

The New Productivity Equation

Old definition: Output ÷ Time = Productivity

New definition: Business Value × Decision Quality × Learning Speed = Productivity

Notice what disappeared from the equation.

Output volume. Time worked. Tasks completed.

Notice what appeared.

Business value. Decision quality. Learning speed.

A team that generates one hundred reports nobody reads has zero productivity by this definition. A team that produces one AI-assisted proposal that closes a $20M customer has extraordinary productivity. The output counts only if it moves the business.

 

---

The Skills That Matter Now

If productivity is no longer about how fast you execute, it is about how well you judge.

The productive people in an AI-era organization:

• Ask better questions than the AI will generate on its own
• Spot confident errors (the AI that sounds right but is wrong)
• Understand when AI output fits the context and when it is generic
• Decide what not to produce (the report nobody needs, the analysis that wastes time)
• Integrate AI output into human strategy without letting the machine thinking override human judgment

These are metacognitive skills: skills about managing cognition rather than performing it.

The old productivity rewarded the fastest executor. The new productivity rewards the sharpest evaluator.

---

What Executives Should Measure Instead

Stop measuring:

• Lines of code or Story Points per developer
• Hours logged per employee
• Reports generated per analyst
• Tickets closed per support agent
• Tasks completed per person

Start measuring:

• Business outcomes per decision
• Strategic insight per analysis cycle
• Customer impact per shipped feature
• Revenue per sales interaction
• Problems solved per team member

The first set is activity. The second set is impact.

Your competitor already started measuring this. Their dashboard looks worse than yours (fewer reports, fewer tickets, fewer hours logged). Their bottom line looks better (more revenue, better customer outcomes, faster decisions).

By the time you notice, they are already two years ahead.

---

The Strategic Shift

The old question was: "How can we make people work faster?"

The new question is: "How can we make better decisions, faster?"

This is not semantic. It inverts the optimization target.

Optimizing for speed gets you burnout, hallucinations, and fake productivity. Optimizing for decision quality gets you competitive advantage.

A company that measures decision velocity (how fast can we move from insight to action) will out-execute a company that measures output volume. Every time.

---

The Real Productivity Gain

Productivity in the age of AI is not about doing more with less.

It is about directing intelligence (human and artificial) toward the right problems.

One person plus AI can accomplish what five people did before, not because that person is five times smarter, but because they are using the machine to extend their leverage. They prompt it, evaluate it, refine it, and integrate it into larger strategy. The output is the machine's. The value is the judgment.

This is profoundly different from productivity 2.0.

Productivity 2.0 was: computers make us faster.

Productivity AI is: we decide better with machines, so we achieve more with less.

---

The Bottom Line

AI has not simply improved productivity.

It has redefined it.

The era of measuring effort is ending. The era of measuring impact has begun.

If your company is still optimizing for output volume, hours worked, or tasks completed, you are measuring the wrong thing. You are creating pressure to do more work faster, which:

• Burns out employees who use AI to work longer
• Floods the organization with low-quality output that requires verification
• Fills the freed time with more meetings, emails, and work instead of actual space to think

The companies winning now measure something different: business value per decision, strategic insight per analysis cycle, customer impact per feature.

They appear less busy. They ship less volume. They work fewer hours.

They are also 5x ahead.

The productivity crisis is not real. The metric is obsolete.

It is time to measure differently.

The New Product Quadrilateral: Geometry Over Feature Lists

 

For most of software's history, product managers operated from a simple assumption: value is built additive. Ship features, watch adoption, ship more. The unit of thinking was the feature - a bounded, deterministic, shippable thing that a team specifies, builds, tests, and ships.

That assumption is collapsing.

When your product's core engine is a LLM - probabilistic, hallucinating, and improving on its own cadence - the feature is no longer a fixed behavior. It is a distribution. The same input can yield different outputs. "The feature" no longer means anything. What matters is the shape of the system that produces the output.

I call this shape the Product Quadrilateral: Human, Model, Workflow, Harness.

Unlike traditional features, which stack additively, these four sides form a closed system where every element constrains and influences the others. Change one side, and you must redesign all three. The strongest AI products aren't the ones with the longest feature lists. They're the ones that have optimized the geometric relationship between these four forces.

---

The Four Sides

Human: Judgment and Delegation

In traditional software, the human is a user - someone who operates the tool. In AI products, the human is increasingly a delegator and supervisor. They set intent, they review outputs, they provide corrections. The human's role has shifted from "click and execute" to "direct and verify."

This changes everything a PM needs to understand:

• What level of skill can verify the output? (A junior analyst won't catch what a senior expert will.)
• How much does the human trust the system?
• Where is the human willing to hand off judgment, and where do they insist on control?
• What information does the human need to feel safe delegating?

Product managers who design for this new human role don't ask "how many clicks does this task require?" They ask "where should the human intervene, and what must they see to make good decisions?"

Model: Capability with No Moat

The model - LLM or specialized AI - is the probabilistic engine. It produces reasoning, generation, classification. But it is no longer an asset you own or fully control. It's a rented substrate with properties that break traditional product thinking:

• It improves without your roadmap (new model versions arrive on external cadences)
• It regresses without warning (model degradation is a real operational problem)
• It has a cost and latency profile that changes constantly
• It has a jagged frontier: superhuman at some tasks, surprisingly brittle at adjacent ones
• It hallucinates with high confidence

Product Managers no longer ask "which model should we use?" as if picking from a catalog. They ask: "What capability does this task demand, and what model economics and latency constraints can we accept?" Then they test ruthlessly to find the inflection point.

Workflow: The Shape of Real Work

This is where most traditional PMs underweight. Workflows are the actual process the product sits inside - a claims review, a code release, a customer support ticket, a hiring decision. The workflow has messy shape: multiple systems of record, human checkpoints, exception paths, regulatory constraints.

AI adds value only when it accelerates or improves a step (or a chain of steps) inside that workflow. A model operating in the abstract is worthless. A model that takes two hours of investigative work down to five minutes - inside a well-designed workflow - is a business problem solver.

The highest-leverage design decision PMs make is often not about the model at all. It's about the workflow redesign. How does the AI-native version of this process differ from bolting a chatbot onto the existing flow? Usually, it differs radically.

Harness: The Reliability Container

The harness is everything that turns raw model capability into dependable product behavior. It's where the real product work now happens:

• Context engineering: what information the model sees, from where, and what it's deliberately not shown
• Tools and permissions: what the model can read, call, change, with what constraints
• Guardrails and policy: limits, approval tiers, escalation rules, sandboxing
• Evaluation and observability: how you know it works, and how you find out when it stops working
• Memory and state: what persists across steps and conversations
• Feedback loops: how corrections become training signals

The most successful AI products aren't using more advanced models than their competitors. They're using better harness engineering. Stronger context retrieval. Tighter tool access patterns. More sophisticated evaluation suites. These are the compounding assets that create defensibility.

---

Why Geometry Matters More Than Feature Lists

A feature roadmap says: "Here are the capabilities we'll add."

A quadrilateral roadmap says: "Here's what outcome we'll reliably accomplish, and which combinations of Human, Model, Workflow, and Harness will produce it."

These are not equivalent. When you design for the quadrilateral, the geometry matters because the sides are not independent:

• Change the human's role (move from reviewer to operator), and the harness must become simpler; fewer toggles, clearer defaults.
• Upgrade the model, and you might reduce harness complexity - fewer guardrails if the model is more reliable - or you might increase workflow complexity if the model is now capable of multi-step reasoning.
• Redesign the workflow, and you might discover the model no longer needs to do the work you asked it to do, or you might discover it needs entirely different context.
• Strengthen the harness, and you might enable higher human autonomy or lower model requirements.

A feature roadmap hides these dependencies. A feature called "AI ticket resolution" says nothing about who reviews it, what happens when it's wrong, what data it's allowed to see, or which model runs it. Those are exactly the decisions that determine whether the feature succeeds or fails.

---

What This Means for Product Managers

Traditional product thinking was: more features, more value.

Quadrilateral thinking is: better shape, more reliability.

Here's what changes in practice:

From "specification" to "evaluation." You no longer write acceptance criteria like "given X, when Y, then Z." You write evaluation sets: representative inputs, a definition of what good looks like, and a threshold. If you can't describe how you'd measure "good" statistically, you haven't specified the product.

From "feature roadmaps" to "outcome roadmaps." Instead of "Q3: AI draft generator, Q4: tone controls, Q1: auto-sending," you ask: "What prevents this outcome from succeeding reliably?" The answer might be: better context retrieval, stricter approval rules, clearer failure handling, stronger evaluation coverage. None of these are features the user sees. All of them are product decisions.

From "user journeys" to "failure stories." How should this system fail gracefully? What does the product do when the model is wrong, the context is thin, or the user is over-trusting? These aren't edge cases; they're core experiences that define whether the product survives.

From "UI optimization" to "delegation design." How does the human express intent? How do they see what the system did and why? How do they intervene? How do they undo? These become the real UX surfaces, not button placements and color contrast.

From "adoption" to "silent failure." The most dangerous failure in AI products is wrong output that looks right and the user accepts it without question. Traditional metrics (DAU, adoption, engagement) are weak signals when the goal is to reduce human effort and increase accuracy. You need to instrument for silent failure and measure intervention rates.

---

The Geometry Is the Moat

When foundation models are commoditized and everyone uses the same APIs, competitive advantage doesn't come from the model. It migrates outward to three things:

1. Workflow embedding: being the system where the work actually happens, with deep integration into how teams operate
2. Harness quality: accumulated engineering on context retrieval, tool design, evaluation, and reliability that is expensive and slow to replicate
3. Proprietary feedback loops: domain-specific training signals and corrections that make the system measurably better over time than the generic version

These three things form a geometric advantage. They compound. A product with poor model but excellent harness and workflow embedding will outlast a product with excellent model but generic harness and weak workflow fit.

---

The New Unit of Value

The feature is not dead, but it is no longer the unit that explains product success.

The new unit is the outcome delivered reliably; reduced from "how many clicks" to "how often does this task get done correctly without rework, and at what cost?"

A strong product team no longer asks: "What features are on the roadmap?"

They ask: "Which side of the quadrilateral is limiting our ability to deliver outcomes? Is it human skill, model capability, workflow fit, or harness engineering?"

The answer determines where the investment goes. And sometimes the answer is: invest in workflow redesign and harness engineering, not in a better model.

That is the shift from feature thinking to geometry thinking.

The products that win the AI era won't be the ones with the most features. They'll be the ones that built the strongest shape.

---

Note: AI is used to polish this post

The Experts Were Wrong About AI. Again.

 

Let’s explore how artificial intelligence has consistently outperformed the predictions made by experts in recent years. While many anticipated that significant AI milestones would take decades to achieve, models have frequently reached these benchmarks in just a year or two.

Key Areas of AI Progress (Underestimated)

• Mathematical Problem Solving: In a 2025 survey, experts estimated a low probability for AI to contribute to solving a Millennium Prize problem by 2027. However, AI reached this benchmark in September 2026.
• Maths Olympiad: Experts previously projected that AI would reach gold-level skills in the International Maths Olympiad by 2030, a milestone that was actually achieved in 2025.
• Theorems and Publications: A 2023 survey of over 2,000 AI researchers predicted it would take 22 years for AI to prove theorems publishable in top mathematics journals. In reality, preprint servers are now being flooded with AI-authored papers, leading to new challenges regarding content responsibility and author comprehension.

 

Moving forward, physics and other theoretical disciplines may not require a new Einstein, but rather a systematic approach to utilizing AI to read and synthesize the vast existing body of scientific literature.

 

• Programming Skills: AI's programming capabilities have far exceeded expert predictions, demonstrating rapid improvement in a short period.
• LifeCode Bench Pro Performance: Experts previously predicted that the best AI score on this difficult coding test would rise to 14% by the end of 2026. However, by May 2026, AI had already achieved a score of 53.8%.
• Task Completion Time: In April and May 2026, experts estimated that by the end of 2026, an AI would be able to reliably complete a coding job that takes a human expert about 3.4 hours. While the survey was still ongoing, a new AI model had already reached a time of 3 hours and 6 minutes

 

This rapid progress suggests that for digital tasks like coding, where institutional and physical hurdles are minimal, AI development is moving much faster than human analysts anticipated.

• Cybersecurity: A significant acceleration in AI's cybersecurity capabilities, marking a clear departure from the conservative timelines previously projected by experts.
• Rapid Capability Surge: In 2023, cybersecurity experts widely predicted that AI would not be capable of autonomously finding and exploiting system vulnerabilities for several years. However, AI reached this milestone in 2024, far ahead of schedule.
• Autonomous Cyber Defense: While there is no single agreed-upon definition, the field is evolving toward agents that can move beyond simple threat detection to engage in active defense measures, such as system hardening and recovery. Ref: CSET (Center for Security and Emerging Technology)
• Modern Cybersecurity Dynamics: Current cybersecurity now involves a race between AI-powered attacks and AI-powered defenses. While AI platforms are highly effective at detecting anomalies at machine speed, they often still require human oversight to distinguish between legitimate system maintenance/deployments and actual malicious activity.

 

Why Experts Were Wrong

• Academic Bias: Many experts surveyed work in academia, an environment characterized by slow, methodical progress. They struggle to account for the velocity of market-driven economies, especially when those industries are supercharged by hundreds of billions of dollars in investment.
• The Disconnect Between Bits and Atoms
• Digital Success: Tasks that rely purely on software such as coding, mathematical proofs, and scientific writing face almost zero friction. These areas have seen rapid breakthroughs because they bypass physical constraints.
• Physical Constraints: Conversely, domains requiring physical labor, massive infrastructure changes, or complex institutional shifts (like building power plants, mining, or changing corporate workflows) face significant, time-consuming hurdles that do not scale at the same speed as digital algorithms; institutional and logistical hurdles that will take decades.

 

Misplaced Predictions (Overestimated)

• Job Automation: expert predictions regarding job automation have been largely incorrect, with the reality of workplace changes falling significantly short of past forecasts
• Overestimated Impact: In 2020, the World Economic Forum projected that nearly 50% of work would be automated by 2025. A follow-up report in 2025 revealed the actual rate of automation was closer to 22%.
• Specific Example: Roles that were expected to be automated early on, such as truckers and taxi drivers, have seen very little disruption so far, contrary to widespread predictions

The transition of physical, labor-intensive industries is much slower than the adoption of software-based AI tools. While digital tasks (like coding or writing) evolve rapidly, changing workflows in physical companies, building infrastructure, and updating supply chains face significant institutional and logistical hurdles that take decades to overcome

• The 'Intelligence Explosion': The predictions of an imminent 'intelligence explosion' (by 2027) is proven unrealistic, given the immense physical infrastructure required to power it.
• The Prediction: Forecast from Leopold Aschenbrenner, which posited an impending intelligence explosion by 2027. This theory suggested that AI would self-improve at an accelerating rate by building massive infrastructure, including hundreds of gigawatt-scale power plants and supercomputers, facilitated by robots designed by the AI itself.
• The Critique: The scenario assumes an overly simplistic feedback loop where AI can easily command physical reality; ignoring the massive logistical, resource, and institutional hurdles involved in building power grids, mining, and large-scale industrial robotics.
• The Reality Check: This "explosion" narrative contracts with the reality of how technical progress actually happens. While software tasks (coding, math) can indeed see near-zero friction progress, physical-world change (the 'atoms' side of the equation) requires decades of labor and coordination, making an instantaneous 'intelligence explosion' that transforms the physical world in a few short years highly unlikely.

 

Future Outlook: The Bifurcated AI Trajectory

I predict a bifurcated future for artificial intelligence, defined by the stark difference between digital-only tasks and physical-world implementation

• Rapid Growth in Software Domains: Areas like mathematics, coding, and scientific research are expected to see exponential progress. Because these tasks occur within software environments and face minimal institutional or physical friction, we are entering a phase where AI will rapidly tackle long-standing theoretical problems perhaps even accelerating the discovery of new science.
• The 'Physical' Bottleneck: In contrast, industries requiring massive infrastructure, energy, and material resources (such as power plant construction, large-scale mining, and industrial robotics) will likely evolve much more slowly. These transitions will take decades, tempering the likelihood of an overnight 'intelligence explosion' that drastically alters the physical world.
• The Future of Human Contribution: Despite rapid AI gains, the speaker argues that progress will remain accessible. The near future may be characterized by a shift where AI tools enable almost anyone to function as a researcher, leading to a period of democratized scientific discovery, even as the ultimate societal impact whether it leads to total job displacement or a post-work utopia remains an open, debated question.

The Evolution of AI Agent Development: From Prompting to Autonomous Systems

 

The evolution of AI agents follows a clear architectural progression. Early applications focused almost exclusively on crafting better prompts. Today's enterprise-grade agents rely on multiple engineering layers that progressively increase intelligence, reliability, and autonomy.

Each new layer addresses specific limitations exposed by the previous one.

Rather than thinking of AI as "better models," it is more useful to think of it as an evolving software stack. Models remain the reasoning engine. The surrounding architecture increasingly determines real-world performance.

Layer 1: Prompt Engineering

Teaching the AI how to think
Prompt engineering represents the first generation of AI application development. Here, intelligence is largely encoded in the instructions given to the model.

Instead of simply asking a question, developers define:

• Roles
• Objectives
• Constraints
• Output formats
• Reasoning style
• Examples

Example:
Instead of: "Answer customer questions."

Better: "You are a senior customer support specialist for an enterprise SaaS company. Answer politely, cite product documentation, never speculate, and escalate billing questions."

The prompt essentially becomes the agent's temporary personality and operating manual.

Common techniques:

• Role prompting
• Few-shot examples
• Chain-of-thought prompting
• Step-by-step reasoning
• Structured outputs (JSON, XML)
• System prompts
• Prompt templates

Strengths:

• Extremely fast to build
• Little infrastructure required
• Excellent for isolated tasks
• Works well for chatbots and assistants

Limitations:
The model only knows what exists inside its context window. It cannot:

• Retrieve fresh information
• Remember previous work
• Use enterprise systems
• Access databases
• Interact with software
• Perform multi-step workflows

Prompt engineering reaches diminishing returns because no matter how sophisticated the prompt becomes, the model remains isolated.

Layer 2: Context Engineering

Teaching the AI what it needs to know

Once developers realized prompts alone were insufficient, focus shifted toward supplying better information rather than better instructions.

This gave rise to Context Engineering.

Instead of embedding every possible fact into a prompt, the agent dynamically gathers the information it needs before reasoning.

Think of the context window as the agent's working memory. Context engineering determines what information enters that memory.

Typical sources include:

• Enterprise documents
• Databases
• CRM systems
• APIs
• Web search
• Code repositories
• Vector databases
• Previous conversations

This is the foundation of Retrieval-Augmented Generation (RAG).

Instead of hallucinating, the agent retrieves.

Example:

A procurement agent receives: "Compare these vendors."

Instead of relying on training data, it automatically gathers:

• Existing contracts
• Vendor performance history
• Pricing
• Compliance reports
• Risk assessments
• External market data

Only after collecting this information does reasoning begin.

Context Engineering Components:

Information Retrieval

• Semantic search
• Keyword search
• Hybrid search

Memory Management

• Short-term memory
• Conversation history
• Long-term memory

Tool Calling

• SQL databases
• ERP systems
• CRM platforms
• Internal APIs

Filtering
Only relevant information is added. Irrelevant context wastes tokens and reduces reasoning quality.

Why Context Engineering Matters:
Better models help. Better context helps far more.

Many enterprise failures blamed on "weak AI" are actually failures in context engineering.

Layer 3: Harness Engineering

Teaching the AI how to execute work

As organizations attempted increasingly complex workflows, another limitation appeared.

Large projects quickly exceeded available context windows.

Imagine asking an agent: "Build an entire ERP implementation plan."

Eventually:

• Earlier decisions disappear
• Objectives are forgotten
• Inconsistencies emerge
• Duplicated work appears

This is known as "context leakage."

The solution is not a larger model. The solution is an external orchestration layer.

This is Harness Engineering.

What is the Harness?
The harness surrounds the LLM. Instead of the model managing everything internally, the harness manages:

• Tasks
• Checkpoints
• Execution state
• Memory
• Retries
• Tool usage
• Workflow progression

The model becomes only one component inside a much larger software system.

Example Workflow:
Suppose an engineering agent is asked: "Upgrade a 200-service microservice platform."

The harness may:

1. Analyze repositories
2. Create dependency graph
3. Prioritize services
4. Generate migration plan
5. Execute updates
6. Run unit tests
7. Run integration tests
8. Review failures
9. Retry failures
10. Produce deployment plan

Each task becomes an independent execution unit. The harness stores results externally. The model never has to remember everything.

Responsibilities of the Harness:

• Task decomposition
• State persistence
• Checkpoint management
• Retry logic
• Tool integration
• Execution monitoring
• Result aggregation

Why the Harness Matters:
Longer workflows become possible. Complex multi-step processes can be orchestrated. The model focuses on reasoning, not bookkeeping.

Layer 4: Loop Engineering

Teaching the AI how to improve itself
Loop engineering represents the next evolution. Instead of relying on human prompts for every stage, this involves creating self-guided scaffolding outside the harness layer.

This allows agents to autonomously trigger tasks, perform maintenance, fix bugs, and verify their own work without constant human intervention.

What Loop Engineering Enables:

• Autonomous verification
• Self-correction
• Iterative improvement
• Bug detection and fixing
• Continuous optimization
• Long-duration autonomy

Example:
An agent is tasked with "Migrate this data pipeline."

In loop engineering, the agent:

1. Executes the migration
2. Automatically verifies results
3. Detects any anomalies
4. Investigates root causes
5. Fixes identified issues
6. Re-verifies
7. Documents the process
8. Escalates only if unresolved

The system operates over hours or days without human intervention.

Key Components:

• Evaluation frameworks
• Feedback loops
• Error detection
• Self-correction mechanisms
• Autonomy guardrails
• Human escalation paths

Why Loop Engineering Matters:
Agents can operate independently over long durations. This enables truly autonomous systems that improve over time rather than simply executing once and stopping.

The Progression of Autonomy

Layer	Focus	Model Role	Human Role	Complexity
Prompt Engineering	Instructions	Execute single query	Write prompts	Low
Context Engineering	Information	Reason with rich data	Manage data sources	Medium
Harness Engineering	Orchestration	Reason and decide	Design workflows	High
Loop Engineering	Learning	Reason, decide, improve	Monitor boundaries	Very High

What This Means for Builders

If you're building Layer 1 systems: Focus on prompt craft and few-shot examples. Understand that you've hit the ceiling on what prompting alone can achieve.

If you're building Layer 2 systems: Focus on information retrieval and context quality. Better context solves more problems than better prompts.

If you're building Layer 3 systems: Focus on orchestration, task decomposition, and external state management. The model is now a component inside a larger system.

If you're building Layer 4 systems: Focus on evaluation frameworks, self-correction mechanisms, and safety boundaries. How does the system know it succeeded? How does it catch and fix its own errors? How do you prevent drift?

The Competitive Advantage

Most organizations are still in Layers 1 and 2. They're writing prompts and managing context.

The organizations building Layers 3 and 4 right now will have a 2-3 year head start on complex, autonomous systems.

The architecture you build today determines what's possible tomorrow.

The Product Requirement Document Is Dying

 

You wrote a 47-page PRD. Half of it was ignored. Your designers misread the success metrics. Your stakeholders debated the screenshots. By the time everyone signed off, the market had moved.

PRDs aren't just struggling. They're becoming obsolete because the fundamental problem they solved has changed. Requirements matter more than ever. But the way we specify them has to shift.

Why PRDs Ever Existed

The PRD was born from constraint.

When a product manager worked in one building, architects in another, and engineers in a third, you needed a translation layer. Something to turn ideas into written specifications that people could interpret and execute.

The PRD answered six questions:

• What are we building?
• Why are we building it?
• Who is it for?
• How should it behave?
• What does done look like?
• What might go wrong?

It was a coordination mechanism. Imperfect, but necessary.

The Game Changes When Machines Execute

AI agents change the rules.

A coding agent can inspect a repository, analyze architecture, write implementation plans, generate code, run tests, and iterate. An agentic product system can research markets, analyze customer feedback, generate designs, create plans, call APIs, and coordinate multiple specialized agents.

The question shifts entirely. A PRD works fine for coordinating humans. But an intelligent system needs something different: it must interpret intent, execute within boundaries, and continuously demonstrate that an outcome is being achieved. That requires a completely different artifact.

Why Static PRDs Fail AI Agents

When an agent encounters a traditional PRD, it faces five fundamental problems:

1. Intent without boundaries. PRDs describe aspirations ("make onboarding frictionless") but not operational boundaries ("onboarding can take 5-7 minutes, not longer"). An agent needs to know what counts as success, what states are valid, what must never happen, which tradeoffs are acceptable.

The next problem is worse.

2. Ambiguity at scale. A human engineer who reads "users can easily reset their password" can ask clarifying questions and infer conventions from the codebase. An agent chooses a plausible interpretation and implements it at scale. If the interpretation was wrong, you discover it in production.

3. Missing context. A PRD sits in a document system while the code, APIs, data models, design system, and deployment rules live elsewhere. An agent cannot safely plan from product intent alone. It needs the relevant architecture, existing conventions, known technical debt, security policies, and prior decisions.

This compounds quickly.

4. Linearity. The classic PRD implies requirements → design → code → testing. Agents work differently. They propose alternatives, build prototypes to resolve uncertainty, discover conflicts with existing APIs, generate tests while writing code, ask other agents to critique their plans, and revise their task graphs based on results. A static document can't keep pace with this.

5. Missing accountability. When a human team implements a feature, responsibility flows through recognizable roles. When agents plan and execute, the question becomes: who authorized the agent to make this decision, on what evidence, under which policy? A PRD rarely records this. An agent operating in production needs an audit trail, not a 30-page narrative.

---

Left shows traditional PRD as a thick narrative document addressed to humans. Right shows modern stack as interconnected layers (intent, behavioral, contract, evaluation, execution) addressed to AI agents. Notice the shift from prose to precision.

---

What Replaces the PRD

The answer is not "a better PRD 2.0."

The answer is a living product execution system composed of:

• A human-readable statement of intent
• A machine-readable behavioral specification
• Explicit context and policy layers for agents
• Executable acceptance tests and evaluation criteria
• A dependency-aware task graph
• Traceable evidence from implementation through production
• Clear approval gates for decisions that humans must make

Think of it as moving from:

"Here's what we want you to build."

To:

"Here's the problem, desired outcome, context, constraints, policies, evidence thresholds, and definition of success. Determine the best way to achieve it, operate within these boundaries, and continuously demonstrate that the outcome is being achieved."

What This Looks Like in Practice

Instead of a 25-page narrative, product teams maintain five layers:

1. Product Intent Layer (human-facing, narrative)

This is how you communicate why. It replaces the opening sections of traditional PRDs:

• The customer or business problem
• Why it matters now
• Target users
• Desired outcome
• Strategic constraints
• Non-goals
• Ethical or regulatory boundaries

This layer stays readable. It's for strategy alignment, not machine execution.

2. Behavioral Specification (machine-readable)

This is how you communicate what. It's precise enough that an agent can plan from it:

• Actors and permissions
• Inputs and outputs
• User-visible states and transitions
• Business rules
• Error behavior and edge cases
• Performance and reliability requirements
• Explicit non-goals

3. Agent Operating Contract

If an agent is executing your product strategy, it needs to know its boundaries. This layer replaces handoff meetings:

• Which repositories the agent can access
• Which tools it can call
• Which files it can modify
• Coding and architectural conventions
• Security and data-handling rules
• When it must ask for approval
• Which sources are authoritative

4. Evaluation Contract

How do you know the agent succeeded? This replaces post-launch review meetings:

• Representative test cases
• Adversarial cases
• Golden outputs or acceptable ranges
• Quality rubrics
• Safety and refusal criteria
• Latency and cost budgets
• Production feedback signals

5. Execution Graph

This is the work breakdown, generated dynamically:

• Epics, capabilities, tasks
• Technical and design dependencies
• Data migrations
• Rollout stages
• Approval gates

The key insight: you're no longer trying to fit everything into one document. Each layer has one job.

---

The complete system showing Product Intent, Behavioral Specification, Agent Operating Contract, Evaluation Contract, and Execution Graph as interconnected layers. Agents interact with the system on the right, feedback loops flow upward. This is the operational structure that replaces a static 30-page document.

---

The Role Changes

The product manager doesn't disappear. The role evolves.

Instead of coordinating execution, PMs become intent architects and control-plane designers.

Your week changes from:

• Writing tickets
• Grooming backlogs
• Clarifying requirements
• Attending status meetings

To:

• Defining strategy and the outcomes that matter
• Analyzing what the data actually says
• Designing and running experiments
• Setting constraints and approval gates
• Watching agent behavior in production
• Reviewing decisions that affect revenue, risk, or customers
• Predicting what might break

The question changes from:

"How do I describe this feature in enough detail that engineering understands?"

To:

"What should the system optimize for, under what constraints, using what evidence, with what level of autonomy?"

Start Now, Not Later

Here's the uncomfortable truth: you don't need to wait for agents to start this shift. Your current human teams would move faster if you stopped writing feature narratives and started writing intent systems. The PRD is holding you back whether you're using agents or not.

This transition doesn't require abandoning PRDs tomorrow. It requires changing what they contain right now.

Stop optimizing for: "Can engineering implement this requirement?"

Start optimizing for: "Can an intelligent system understand the objective, constraints, evidence, and success criteria?"

Start adding:

• Explicit outcomes (what measurable result?)
• Explicit constraints (what must not happen?)
• Explicit decision rights (what can an agent decide?)
• Explicit evaluation criteria (how will we judge?)
• Explicit context (what should an agent consult?)
• Explicit escalation rules (when must humans intervene?)
• Explicit telemetry (what evidence proves it's working?)

This gradually transforms a PRD into an agent-ready specification.

The Final Shift

The fundamental change is this: execution is becoming autonomous.

When humans do the work, you need detailed instructions. When agents do the work, you need clear intent, relevant context, hard constraints, evaluation mechanisms, governance, and feedback loops. The PRD was built for a world of human execution. We need new artifacts for a world of autonomous execution.

The product manager's job shifts upward in the abstraction stack.

You stop saying "build this" and start saying "achieve this." You stop describing implementations and start defining boundaries. You stop writing acceptance tests and start designing evaluation mechanisms.

The PRD's real value wasn't the document itself. It was reducing ambiguity between teams. Now that teams include machines, ambiguity creates production risk. An agent will confidently execute an ambiguous requirement in ways that break your product.

The product manager's real job becomes designing the system in which intelligent agents make good decisions.

What this means for you next week:

Pick one upcoming feature. Instead of writing a 15-page PRD, write:

1. A 300-word problem statement (human-readable intent)
2. Five concrete success criteria (behavioral spec)
3. A one-page "what must never happen" list (boundaries)
4. How you'll know it's working in production (evaluation)

That's your starter template. No agents required. Your current team will ship faster.

The successor to the PRD isn't just a better document. It's a system of durable intent, clear boundaries, and continuous evaluation. Start building it now not when you have agents, but when you have the next feature to ship.