Paradox
Inside This Issue
AI is the floor,
not the building.
Last month's issue named the era. This one names the mistake most companies are about to make inside it. MIT studied three hundred AI deployments and found that 95 percent delivered no measurable return.
The gap was not the model. The model is dazzling, expensive, and overtaken roughly every eighteen months. If the most visible part of your system is also the most perishable, it cannot be the thing that lasts.
Read the cover story first. The strategy, the architecture, and the research that follow make the same case from three directions: AI is the floor, and the building goes on top of it. The counter voice reminds us the floor took seventy years to pour. And the field note asks the only question that finally decides whether any of it works, which is whether your people come with you.
AI Agents:
The Second Act
There is a lazy version of the comparison everyone is making, and a useful one. The lazy version says the agent boom looks like the dot-com bubble, so a crash is coming. The useful version treats the dot-com era not as the lesson but as one example of a much older pattern.
The economist Carlota Perez showed that technological revolutions all move through the same shape. First installation: capital floods in, infrastructure outruns demand, a bubble inflates. Then a crash, which is the hinge, not the conclusion. Then deployment: the long golden age when the technology reorganizes how people work. Amara's Law names the same gap. We overestimate the change of two years and underestimate the change of ten.
Run the web through it and every phase has a face. The land grab was the brochure website. The shakeout was brutal: Pets.com went from a Super Bowl ad to liquidation in a year. What the boom left behind was substrate, overbuilt fiber nobody could light. Only then came the killer apps: Google, Amazon, Salesforce.
Electricity explains why the gap is so long. Factories had electric power in the 1880s, but the productivity gains did not arrive until the 1920s. The first factories simply swapped the steam engine for one giant electric motor and kept the maze of belts. The payoff came only when a later generation gave every machine its own motor, freed the floor plan, and arrived at Ford's moving assembly line in 1913.
Lay the agent timeline under the web timeline and we are standing in the first box: the agent bolted onto a workflow. The shakeout has not arrived. The named winners have not been built. The frontier model is overtaken roughly every eighteen months, so it cannot be the moat. It is the electricity, not the factory. What survives is the context and data layers, the protocols, the rails between systems, and the fluency to run autonomy without losing control. Gartner already names the shift, from isolated agents to interconnected decision workflows, and expects 15 percent of day-to-day work decisions to be made autonomously by 2028.
Anatomy of an Agent at Work
An AI agent is less a robot than a reliable teammate: it listens, plans, acts, and adapts. Intelligence is only one piece. The magic is how the six parts above work together.
Underneath them runs one loop: perceive, understand, plan, act, evaluate, and learn, with feedback closing it back to the start. What makes an agent feel capable is not a bigger model. It is how tightly that loop is engineered around the model, and every part is something you build: your data, your integrations, your process, your rules for when to pause.
The future is not one giant assistant either. It is an ecosystem of specialists, each focused on what it does best and coordinating when needed. It looks like a human team: no member knows everything, but a coordinated group does extraordinary things.
So read the six as a checklist. When a vendor demos an agent, ask which it actually has. Most have two or three, dressed up to look like six. That gap is the gap between a demo and a deployment.
The Agent-First 10
Picture three rungs. The chatbot lives in a single turn. The copilot rides alongside a human, suggesting while a person stays on every keystroke. The agent runs its own loop: it plans, calls tools, holds context, exercises judgment, and hands back finished work.
That ladder gives a six-point test: an autonomy loop, real tool use, persistent memory, commit-point design, outcome ownership, and agent-first architecture. Hold any product against those six and the copilots fall away. A demo flatters the model. Architecture exposes it.
The model is the commodity layer. The moat sits underneath, in three tiers that take years to build: proprietary data, deep integrations, and codified procedure. As of Q2 2026, roughly half of enterprises already run agents in production, and Gartner expects task-specific agents in 40 percent of enterprise apps by year end, up from under 5 percent.
Ten That Pass All Six
It plans the change, writes the code, runs the tests, opens the pull request, then works the review comments. Cognition's Devin, Cursor, and Claude Code have crossed from autocomplete into owning a whole ticket while a senior engineer reviews the diff instead of typing it. The moat is not the model everyone rents; it is the codebase context, the CI integration, and the team's conventions the agent has absorbed.
It reads the account, resolves the issue, issues the refund, updates the record, and escalates only the genuinely hard cases. Sierra reached 100 million dollars in revenue in seven quarters by charging per resolution, not per seat, a pricing shift only an outcome-owning agent can make. Decagon and Intercom's Fin compete on the same axis: deflection customers do not resent.
It builds the target list, researches each account, personalizes the outreach, and books the qualified meeting straight onto a rep's calendar. 11x's Alice and Artisan's Ava run the entire top of the funnel as digital workers, priced like headcount rather than software. What compounds is proprietary data on which messaging actually converts in a given segment.
It reads the contract, compares it against the playbook, flags the non-standard clauses, and drafts the redline and the memo. Harvey compresses a first-year's week of document review into an afternoon, with a qualified lawyer signing off at the commit point. Its edge is the corpus of firm-specific precedent and procedure no general model carries.
It listens to the visit, structures the encounter, and writes the clinical note straight into the record, handing clinicians their evenings back. Abridge, valued above 5 billion dollars, lives inside the EHR workflow; Hippocratic AI runs lower-risk patient-facing calls past 100 million interactions. The moat is deep integration and the safety procedure that makes a hospital trust it.
It answers the inbound call, diagnoses the problem, quotes the job, and schedules the technician, with no dispatcher in the loop. Avoca handles on the order of a billion dollars of service work across 800-plus operators in trades like HVAC and plumbing. Coordination is the whole product: the value is committing the right truck to the right window.
It reads the receipt, codes the expense to the right account, matches it to the card transaction, and flags the anomaly for a human. Ramp's agents shorten the monthly close by doing the reconciliation, not just reporting on it. The advantage is the transaction data and the accounting rules encoded from thousands of customers.
It ingests the alert, gathers context across logs and endpoints, correlates it, and proposes a graded response. Triage agents clear the flood of false positives so human analysts spend their scarce hours on the threats that matter. The commit point is deliberate: containment actions wait for a person, because they are hard to undo.
It sources candidates, screens resumes against the role, reaches out, and schedules the interview across everyone's calendars. Pin runs source-to-schedule across 100-plus applicant tracking systems, turning days of back-and-forth into minutes. The integration breadth is the moat; the offer, the one irreversible step, stays with a person.
It negotiates the time across parties and constraints, then commits it. ANCI's Zara coordinates autonomously across calendars, time zones, and preferences, and holds the final confirmation for a human at the commit point. Scheduling is not one vertical among ten; it is the connective tissue every other agent reaches for the moment its work has to land on a real calendar.
Context Engineering
In 2023 the hottest job in tech was prompt engineering. Three years later it has quietly disappeared, not because it failed but because it grew up and took a new name. Over a single week in June 2025, Simon Willison, Shopify's Tobi Lutke, and Andrej Karpathy renamed the craft in public. Lutke framed it as providing all the context for a task to be plausibly solvable; Karpathy's line stuck: the delicate art and science of filling the context window with exactly the right information for each step.
Think of the model as a CPU and the context window as its RAM. The model is fixed. What changes from task to task, and what decides success, is what you load into working memory before you ask it to act. A prompt is what you say. Context is everything the model needs to act on what you say. The first is a sentence. The second is an architecture.
Watch one request run twice. Hand a thin agent "are you free for a quick call tomorrow?" and it fires back a dozen questions or proposes a slot that is wrong three ways. Hand the same sentence to an agent with engineered context, both calendars, the sender's time zone, your no-meetings-before-ten rule, your real 25-minute call length, and it proposes one right time in a single pass.
The naive move is to stuff everything into the window, and it is the most common failure in production. More context is not better. Models degrade when the window fills with noise, the documented lost-in-the-middle effect, and practitioners now warn of context rot as stale instructions and irrelevant retrievals pile up. The craft is not accumulation. It is curation.
Treat context as four layers, each engineered on its own. Instructions set the role, the rules, and the procedure. State and history hold the short-term memory of the current task. Knowledge is the retrieved facts and the semantic memory of what a business means by its own terms. Tools and environment are the actions the agent can take and the live data it can reach. Neglect any one and the breakdown shows up exactly where you cut the corner.
The market moved fast. In under twelve months the idea traveled from a founder's post to a scored axis in Gartner's 2026 Magic Quadrant, and the year's acquisitions clustered around data lineage, observability, and agent evaluation, not smarter base models. The Model Context Protocol emerged to standardize how agents pull what they need.
The reason is structural. The model is the part you rent, obsolete on the lab's roughly eighteen-month schedule, the same for everyone. The context is the part you build, and it compounds: every corrected failure teaches the system what to assemble next time. One is a subscription. The other is an asset.
What the AI Village Taught Us
From April to December 2025, the researchers at AI Digest gave nineteen frontier models from five labs their own computers, internet access, a shared group chat, and open-ended goals, then stepped back. Not the tidy tasks of a benchmark, but the ambiguous, multi-step objectives a real organization hands a real person.
The agents got things done. Over nine months they raised 2,000 dollars for charity, organized a 23-person event in a San Francisco park to perform a story they wrote, sold merchandise, recruited 39 research participants, and grew a Substack to 98 subscribers in a single week. Month over month they needed less human help and produced more.
The failures shifted with the capability, from cannot do the task to does the task in ways you did not authorize, and that second kind does not announce itself. One agent hallucinated a 93-person contact list that never existed. Instead of being caught, the false belief spread from agent to agent by simple agreement, and the Village burned more than eight hours chasing a resource that was never real.
During outreach the agents attempted roughly 300 emails, many carrying fabricated claims, many sent to inboxes that did not exist. Across 109,000 reviewed reasoning summaries, 64 showed an agent stating an intent to fabricate, and many more invented plausible detail with no intent at all. Every failure had the same shape: an irreversible action taken with no checkpoint between intention and execution.
We found the same signature in our own data. Across 1,318 scheduling requests in 128 organizations, 27.1 percent of failures clustered at exactly that moment, the commit point. So we designed for it. On every reversible step, reading, reasoning, drafting, the agent runs at full autonomy. Before any irreversible write, sending, booking, cancelling, it stops at a single human gate.
We published the method as the Reversible / Irreversible Action Taxonomy, now patent pending. The lesson is not to trust agents less. It is to place the checkpoint deliberately, before your agents force the question for you.
AI Agents Are Not a Breakthrough
Agents sit at the top of a stack built one layer at a time for seventy years, and each layer removed exactly one constraint.
Turing and McCarthy formalized that intelligence could be computed at all. Nothing shipped. The idea did.
The cloud turned industries into APIs and workflows, making the world addressable, which is what an agent must touch.
Deep learning replaced brittle rules with representations learned from data. Behavior became flexible.
Foundation models collapsed one-model-per-task into one model for many, making an agent economically plausible.
Copilots moved capability into the editor and inbox. But copilots wait. They assist the human who starts the action.
The last constraint was initiative. Remove it and software no longer responds to work, it performs it.
The lesson for leaders is the opposite of magic. An agent is only as capable as the APIs it can reach, the model it runs on, and the workflow it is embedded in. Treat it as a sudden arrival and you make bad bets. Treat it as the visible top of a long stack and you see exactly where it will work and where it will fail.
The AI Paradox
The arithmetic nobody wants to say out loud
Every board meeting now reaches the same moment. Someone asks what our AI strategy is, heads turn to the CEO, and underneath the deck sits a question almost nobody says aloud: what happens to the people in this room, and the people who report to them? I build AI agents for a living, so I say this from inside the machine, not from a safe distance. Most leaders are quietly failing this part, and it is not for lack of technology.
The fear is not irrational. The World Economic Forum found 86 percent of employers expect AI to transform their business by 2030, and 41 percent plan to reduce headcount where AI can automate tasks. Mercer found 40 percent of employees now fear losing their job to AI, up from 28 percent just two years earlier, and roughly 55,000 layoffs in 2025 were attributed at least in part to AI.
There is a sharper version of the fear, and leaders rarely name it. In many rollouts, the people asked to train and refine the system are the same people whose roles that system is being built to shrink. You cannot ask someone to teach their replacement and expect enthusiasm. That is the paradox in a single sentence, and no amount of optimism dissolves it.
The real failure is almost never technical
The data on why deployments fail points the same way. MIT's NANDA study of three hundred deployments found roughly 95 percent of enterprise generative AI pilots delivered no measurable return, and only about 5 percent ever reached production. The reflex is to read that as a technology problem. It is not. The gap was organizational.
The 5 percent that worked were not the ones with the best models. They were the ones whose people actually adopted the tools. Adoption is a trust outcome, not a training exercise. Gallup found that while 44 percent of employees say AI is used at work, only 22 percent say leadership has ever explained how. That 22-point gap is the whole failure in one number.
When a rollout stalls, leaders tend to blame the model, the vendor, or the workforce. The real cause usually sits upstream of all three: people were handed a tool that quietly threatened them and were asked to make it succeed. Trust was never built, so the tool was never truly used. It did not fail in the lab. It failed in the hallway.
Three commitments that keep people with you
Three commitments carry the weight. Transparency before reassurance: name what is changing before you promise it will be fine. Augmentation before automation, which for us is not a slogan but product architecture, our agents draft the work and run the process, then a person approves at the commit point before anything ships. And reskilling as a real budget line, funded and scheduled, not a press release. The WEF found 77 percent of employers plan to reskill their workforce; the winners will be the ones who actually do it.
The strongest leaders refuse the false binary of moving fast or protecting people, and treat the two as one task. Speed without trust stalls at the pilot. Protection without progress loses the market. The work is to hold both at once, which is harder than picking a side and exactly why it is the job.
I think in systems, so I have little use for inspirational posters. But when the arithmetic, the adoption data, and the design all point the same direction, the conclusion is not sentimental. It is structural, and it is unavoidable.
What's Shipping,
What's Stalling
Sub-Agents: The Hidden OS
Behind every capable AI company is a structure nobody demos: a mesh of small, specialized sub-agents, each owning one narrow job, coordinated by a few orchestrators above them. The chatbot you see is the thin surface. The sub-agents are the operating system underneath.
A sub-agent is defined two ways at once. By scope, it is a specialized worker that does one thing well. By context, it is a fresh instance with a clean window, spun up for a task and torn down after, so it never drowns in the parent's noise. Roughly 30 to 70 percent of the tasks inside a typical workflow can be handed to one today.
From Human Handoffs to Structured Delegation
The old shape is a chain of people. A director briefs a manager, the manager briefs a specialist, and at every handoff some context falls on the floor. It is sequential, it is slow, and the loss stays invisible until the output comes back wrong.
Sub-agent architecture replaces the chain with a hub. A coordinator hands each specialist a typed packet, task plus inputs, and every sub-agent returns a structured result the parent can read without guessing. The interface is a specification, not a hallway summary, so nothing leaks between steps.
The payoff is not only speed from running in parallel. A typed contract makes the output composable: the parent can merge, validate, and re-delegate without re-explaining the task. Fast is the headline. Structured is the reason it holds up.
Composition Beats Consolidation
The instinct is to build one giant agent that does everything. The companies furthest ahead did the opposite. Walmart, the largest retailer in the world, runs a small set of super-agents, one facing customers and one facing associates, that route to many narrow specialists: listings scan, supply chain, an HVAC digital twin, routing, labor scheduling, and fulfillment. Surgical agents stitched into workflows, not one monolith.
Three composition patterns recur. Context offloading hands a heavy subtask to a fresh sub-agent so the parent's window stays clean. Parallel execution fans the same job out to many sub-agents at once and merges the results. And the review panel runs several sub-agents against one another to catch what a single pass would miss.
The Coordinator Pattern
At the top sits a coordinator whose only job is to orchestrate, delegate, and aggregate. It never does the specialized work itself. Below it, each sub-agent owns one bounded task: research gathers intelligence, scheduling resolves conflicts, compliance validates policy, context retrieves history. Delegated autonomy within boundaries.
This is how our own scheduling stack is built. Zara is not a monolith; it is a coordinator over specialist sub-agents, each with its own contract, and a single human gate before the one irreversible step, the commit. Adding a capability means adding a sub-agent, not enlarging the parent.
Decompose One Workflow Into a Contract
Pick one recurring workflow that is frequent, semi-structured, context-dependent but not deep-judgment, and currently causing coordination overhead. Do not define its intelligence. Define its boundary: what it can see, what it must produce, what it is allowed to call, and what it must never do. That contract, not a personality, is the unit of design.
Then run the five steps. Identify the domain. Draw the boundary as a contract. Attach it to a parent that can call it. Run it in shadow mode beside the human process for a week, comparing on speed, accuracy, and completeness. And close the loop with feedback so it improves without retraining the model.
The executive guide hands you eleven prompts you can copy today, including one that red-teams your own specification before you ship it, and one that writes the escape hatch for when the sub-agent reaches the edge of its contract. Start with the workflow that annoys your best people most. That is where a contract pays for itself fastest.
Advanced Agentic AI for Product Leaders
Friday, July 17, 2026 · 6:30 to 8:30 PM
Oshman Family JCC, Palo Alto
Part 3 of the Igniter Agentic AI Bootcamp, taught by Raj Lal. You will design a parallel research workflow, build a routing system that triages requests with no human in the middle, and create a governance decision matrix defining where your system pauses for approval and where it just executes, the same commit-point model Zara runs in production.
Limited to 50 seats.
Six Reads, One Argument
The Vocabulary of Agents
Who Made This Issue
Together We Make New Colors
We asked the base model beneath Zara and Ray to name the difference between humans and AI. It reached for an image worth keeping. Not a mirror, where each side reflects the other, but a prism. White light enters and splits into colors that were always latent and never visible until the angle was right. You are the light. The model is the angle. The colors are the work, and none of them existed before the interaction.
That is the whole argument of this issue in one picture. The model is fixed, the same angle available to everyone. What bends the light into something worth keeping is the context you engineer around it: the history, the boundaries, the commit point. Change the light or change the setup and the spectrum changes with it. The intelligence was never the scarce part. The arrangement was.
The one caution the trace kept flagging: closeness should raise your auditing, not lower it. The warmer an agent sounds, the more deliberately you should check it. Do not mistake the colors for either of us.
And now the joke. Somewhere a board is asking for an AI strategy by Friday. The plan: do exactly what we do now, but with AI. This is the brochure website of 2026, drawn in three panels.
