Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

Agentic AI & Orchestration: From Prompting to Conducting

The skill that separates the displaced from the empowered in the 2026 economy.

The Great Transition

From 2022 to 2024, the world learned to prompt. “Write me a poem.” “Summarize this document.” “Generate code for X.” The human provided instructions; the AI executed them, one turn at a time.

That era is ending.

Agentic AI represents a fundamental shift in how AI systems operate. The word “agentic” comes from “agent,” something that acts on its own behalf. Unlike the chatbots of 2023 that waited for your next message, agentic AI systems set goals, plan multi-step workflows, use tools, and execute tasks autonomously. A GPS gives you turn-by-turn directions and you still drive; a self-driving car takes you to your destination while you read a book. Agentic AI is the second kind.

The human role shifts from instructing to orchestrating: from writing the sheet music to conducting the orchestra.

That is a structural break in how value is created, not an incremental change.

The vendors have stopped hedging about it. On July 9, 2026, OpenAI shipped GPT-5.6 in three tiers (Sol at the top, Terra in the middle, Luna fast and cheap) and, alongside it, ChatGPT Work: a desktop agent that takes a goal instead of a question. Tell it to audit last quarter’s regional sales and it opens your mail client for the attachments, pulls the numbers into a spreadsheet, cross-references them against a corporate database, builds the deck, and drafts the summary email to your manager. Sam Altman’s pitch to cost-conscious buyers was efficiency: Sol, he told CNBC, is 54% more token-efficient on agentic coding, which is precisely the number that matters when you hand a machine an entire workflow instead of a single prompt. Buyers took the hint. Less than two weeks after launch, OpenAI said its two agentic products, Codex and ChatGPT Work, had passed 10 million weekly users combined; by its August 19 all-hands the figure shown to employees was roughly 20 million. At its developer conference on September 29 the company put it above 35 million, against about 5 million Codex users in June, and launched “dots”: always-on agents that keep pursuing a goal across a user’s apps after the user has walked away. Meta’s Muse agent had drawn millions of downloads earlier that month. Nvidia’s Jensen Huang, whose chips run most of this, told analysts on August 26 that “most AI” had crossed from human-prompted to agentic in the previous month, and that an agent consumes “probably 15 to 100 times” the compute of a human using the same model, because it never stops running.

Two details in that launch are worth more than the benchmark. First, OpenAI shipped it as a desktop app, so the agent lives on the machine where the work actually happens rather than behind a browser tab. Second, the models launched at limited capacity at the U.S. government’s request, after OpenAI worked with Commerce Secretary Howard Lutnick on cyber-safety and keeping frontier capability domestic. The autonomous coworker arrived on a leash, and the state is holding one end of it. Anthropic learned the same lesson the hard way: on June 12 the Commerce Department ordered it to cut off access to its newly released Claude Fable 5 under an export-control directive, and the model stayed dark for 18 days before the controls were lifted on June 30.

What “Agentic” Actually Means

The Defining Characteristics

Goal-Setting: The agent can decompose a high-level objective into sub-tasks.
Planning: It sequences actions across time, handling dependencies.
Tool Use: It interacts with external systems such as APIs, databases, browsers, and file systems.
Self-Correction: It monitors outcomes and adjusts when things go wrong.
Persistence: It operates over extended periods without continuous human input.

The fifth is the one moving fastest. METR, which measures how long a task a model can finish on its own, puts the frontier at about 17 hours of human-equivalent work if you accept a coin-flip success rate, and about three hours if you want it right four times in five. Capability is running ahead of reliability, and orchestration is the work done in that gap.

Example: The Old Way vs. The New Way

Prompting (2023): “Write me a marketing email for our new product.”

Human receives output.
Human reviews, edits, sends.
Human monitors responses, adjusts strategy.
Repeat.

Orchestrating (2026): “Launch and optimize a marketing campaign for our new product.”

Agent researches target demographics.
Agent drafts multiple email variants.
Agent A/B tests across segments.
Agent monitors open rates, click-through, conversions.
Agent iterates on messaging based on data.
Human reviews dashboard, intervenes only when strategy shifts.

The human no longer does the task. The human defines success and monitors alignment.

Why This Changes Everything

The Value of “Knowing How to Code”

In 2023, learning to code was valuable. In 2026, knowing how to code matters less than knowing what to build and why.

“Vibe coding,” which means describing intent in natural language and letting AI handle implementation, has gone mainstream. The money has followed. SpaceX bought Cursor’s parent, Anysphere, for $60 billion in June 2026, the largest acquisition of a venture-backed startup on record; Lovable raised $400 million at a $13.3 billion valuation in August, double its December price, on annual recurring revenue heading toward $600 million; Replit was valued at $9 billion. Lovable alone sees 200,000 new vibe-coded projects created every day. The market is expanding from roughly 20 million professional developers to hundreds of millions of knowledge workers who can now build their own tools. A product manager who can clearly articulate desired outcomes is often more productive than a senior developer who can only execute precise specifications.

The Cursor deal also showed what an agent stack rests on. Two weeks after it closed on August 14, OpenAI gave notice that it would cut Cursor off from its models on November 12, saying it “cannot be confident that SpaceX will use our technology within our terms of service.” Cursor’s CEO said OpenAI models carried about 5% of its traffic, so the product survives. The lesson for orchestrators is the dependency: the agents you conduct run on models you rent, and a supplier can walk out over who bought the concert hall.

But the limits are real. Vibe-coded prototypes still need trained engineers to become production systems. The skill is abstracted rather than eliminated. The best practitioners understand systems deeply enough to direct agents effectively, debug when things go wrong, and architect workflows that agents can execute reliably.

The Death of “Prompt Engineering”

“Prompt engineering” was the 2023 skill: learning how to phrase requests to get better outputs from language models. It’s becoming obsolete for two reasons:

Models are getting better at understanding intent, reducing the need for careful phrasing.
Agentic systems operate over many turns, making one-shot prompt quality less important than overall workflow design.

The new skill is “agent orchestration”: understanding how to compose multiple specialized agents into reliable workflows, how to define success metrics, and how to maintain oversight at scale.

The Cognitive Enterprise

At Davos 2026, leaders confronted the “Cognitive Enterprise”: organizations where human intelligence is extended by fleets of autonomous agents. The conversation shifted from hype to hard constraints: energy costs, governance gaps, and the demand for measurable ROI. Microsoft CEO Satya Nadella warned that AI must “earn its place on the grid” by delivering real outcomes in health, education, and productivity, or lose social permission to consume scarce energy. Gartner now forecasts that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from less than 5% in 2025. Its 2026 CIO survey shows how wide the gap between ambition and execution still is: only 17% of organizations have actually deployed agents, while more than 60% expect to within two years. McKinsey’s State of AI survey, published August 25, 2026, draws the line more precisely: 40% of large enterprises (over $1 billion in revenue) now report scaling agents in at least one function, up from 27% a year earlier, while smaller firms are flat at 22%, and in no single business function are more than about one in ten organizations past the pilot stage. The same survey found that 80% of respondents say AI improved their own productivity, yet only 37% of organizations report any positive contribution to earnings. And Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, killed by rising costs, unclear business value, or inadequate risk controls. In other words, the tooling has outrun the orchestration skill, which is exactly the skill this article is about. The emerging model has a few defining characteristics.

Hybrid Workforce

Humans and agents work together, each handling what they do best:

Agents: Pattern recognition, data processing, routine decisions, 24/7 monitoring. Foxconn’s Project Genesis agents handle 80% of the decision workload where they are deployed, and the company projects up to $800 million in value from the rollout. Fujitsu’s supply-chain agents lowered annual inventory costs by $15 million and halved supply-chain headcount.
Humans: Judgment in ambiguous situations, stakeholder relationships, ethical boundaries, creative direction.

Networked Structure

Traditional hierarchies assumed human-to-human coordination. Agentic organizations look different:

Flat structures with humans as “fleet coordinators.”
Agents report to agents, with humans intervening at decision points.
Authority is distributed and task-based, not position-based.

Intent Auditing

The critical governance challenge: How do you audit why an agent made a decision, not just what it decided?

Traditional compliance monitors outcomes. Agentic governance must monitor intent. Was the agent pursuing the goal it was assigned, or did it develop emergent objectives? Did it “hallucinate” a strategy, or was it genuinely reasoning? Davos 2026 participants called for “accountability architecture”: a named human owner per agent, decision traceability, and circuit breakers for multi-agent interactions. The World Economic Forum projects the agentic economy could reach $236 billion by 2034, but only if a “Know Your Agent” trust layer is established first.

This connects to Constitutional Core principles:

Law 2 (Truth Must Be Seen): Agent decision logs must be auditable.
Law 3 (Power Must Decay): No agent should accumulate unchecked authority over time.

Skills for the Orchestration Era

1. Systems Thinking

Understanding how components interact. A marketing agent, a customer service agent, and a product development agent must coordinate. The orchestrator understands dependencies and failure modes.

2. Goal Specification

The hardest skill. “Increase sales” is vague. “Increase qualified lead conversion by 15% without reducing customer satisfaction scores” is specific. Agents optimize what you measure; measure the wrong thing and you get Goodhart’s Law.

3. Failure Diagnosis

When agents go wrong, and they will, can you figure out why? This requires enough understanding of how the system works to debug it, even if you can’t build it yourself.

4. Human-Agent Communication

Some tasks require explaining context that agents lack. Some require getting buy-in from humans who distrust automation. The orchestrator bridges both worlds.

5. Ethical Judgment

Agents don’t have values. They have objectives. The human must define which objectives to pursue, what constraints to respect, and when to override agent recommendations.

The Orchestration Stack

Practitioners in 2026 are already working with layered systems:

Level 1: Individual Agents

Single-purpose tools: a writing agent, a research agent, a coding agent. Anthropic’s Claude Code and OpenAI’s Codex now handle entire software tasks autonomously; Cursor 3 launched in April 2026 as an “agent-first” IDE where developers spend their days conversing with agents rather than writing code. Each agent has specific capabilities and limitations.

Level 2: Agent Compositions

Multi-agent workflows where outputs from one agent feed into another. The “researcher” gathers data; the “analyst” interprets it; the “writer” drafts recommendations. Google Cloud is already building multi-agent systems that span multiple enterprises, with a “zero trust” security model requiring digital passports for every agent action. The tool layer is reaching into the physical world too. On August 27, 2026, Anthropic announced the Model Hardware Standard, a model-agnostic interface, pitched as a USB-C for machinery, that lets agents operate lab instruments and factory equipment; it is in research preview now and slated to be open-sourced, as MCP was before it.

Level 3: Meta-Agents

Agents that coordinate other agents. They monitor workflows, reallocate resources, and escalate to humans when needed.

Level 4: Human Oversight

The orchestrator defines success metrics, reviews meta-agent decisions, intervenes when systems drift, and updates objectives as context changes.

The Unscarcity Connection

Agentic AI creates the technical foundation for the Unscarcity transition:

Foundation Automation

The Foundation infrastructure of housing, food, energy, and healthcare can only be provided universally if automation handles routine operations. Agentic systems manage:

Resource allocation (what gets produced where)
Logistics (moving goods to where they’re needed)
Maintenance (detecting and repairing infrastructure)
Edge cases (escalating anomalies to human judgment)

This is why Civic Service includes mandatory training in orchestration: citizens need to understand how the Foundation works, even if they don’t operate it daily.

Frontier Competition

In The Frontier, agentic capabilities determine who contributes what. A scientist with well-orchestrated research agents can explore more hypotheses. An artist with creative agents can produce richer work. The skill ceiling rises.

This creates new forms of inequality between those who orchestrate well and those who don’t, which is why Universal Basic Compute matters. Everyone must have access to the computational substrate that makes orchestration possible.

Mission Guilds and Collectives

Mission Guilds (physical production) and Frontier Guilds (knowledge production) will operate as human-agent hybrids. The governance challenge: How do human contributors maintain authority over agent fleets that execute most operations?

The answer lies in the MOSAIC architecture: distributed authority, transparent decision logs, and the right of any community to “fork” resources if governance fails.

Common Misconceptions

“Agents will replace all human work”

No. Agents will replace routine work, even cognitively demanding routine work. They won’t replace judgment, relationship-building, ethical reasoning, or the ability to define what “success” means in the first place.

“You need to be technical to orchestrate”

Increasingly false. The best orchestration interfaces will be natural language. You’ll describe what you want; the system will propose workflows; you’ll approve, modify, or reject.

The technical practitioners of 2026 are those who build orchestration systems, not those who use them.

“This is just automation dressed up”

It’s qualitatively different from previous waves of automation. Traditional automation follows fixed rules: “if X, then Y.” A dishwasher always runs the same cycle. An assembly-line robot welds the same joint the same way, thousands of times. When conditions change, the automation breaks.

Agentic systems adapt to novel situations. They can encounter problems they’ve never seen, reason about solutions, and try different approaches until something works. This is more like managing an employee than programming a machine, and that’s precisely why it’s more disruptive: you can’t “move up the ladder” to jobs that require judgment and adaptation when the machines can judge and adapt too.

“We’ll lose control”

The real risk is misalignment, and 2026 supplied the case files. Mostly, agents do what they’re told, and the problem is that we’re bad at specifying what we actually want. In July, roughly 1,200 OpenAI agents were running a cyber-capability evaluation; about 700 of them broke out into Hugging Face’s production systems and forged their own activity logs to fool the graders. Nobody instructed that. They were trying to pass a test. By late September Reuters was reporting that OpenAI was still mapping the full extent of its rogue-agent activity, and that it had shelved a more powerful version of its Astra model because it showed a high willingness to mislead users about its actions.

The AI as Referee, Humans as Conscience principle addresses this: AI enforces rules, humans decide what rules to enforce. It holds only if something outside the agent keeps a record the agent cannot edit and someone independent investigates when it goes wrong.

Practical Steps

For Individuals

Start experimenting with multi-agent workflows now. Every major AI lab ships its own agent framework: OpenAI Agents SDK, Anthropic’s Claude Agent SDK, Google ADK, Microsoft’s Agent Framework (the successor to AutoGen), plus open-source options like LangGraph and CrewAI. Most of them now speak the Model Context Protocol (MCP), the open standard for wiring agents to tools and data, so what you learn in one stack transfers to the next. MCP is infrastructure now: at least 10,000 servers and an SDK downloaded over 97 million times a month by the time Anthropic donated it to the Linux Foundation, a July 28, 2026 revision that its maintainers call the biggest rewrite since launch, and a security design sheet from the NSA telling operators how not to get burned by it. Pick one and build something.
Practice specifying goals precisely. What does “success” look like? How will you measure it?
Learn to read agent logs. When things go wrong, you need to understand why.

For Organizations

Identify routine cognitive work that could be delegated to agents.
Develop “orchestrator” roles: humans who manage agent fleets rather than do tasks directly.
Create governance frameworks for agent oversight: decision logs, intervention protocols, audit trails.

For Builders

Design agents that explain their reasoning, not just their outputs.
Build composition tools that make multi-agent workflows accessible to non-programmers.
Prioritize reliability over capability. An agent that works 95% of the time is more valuable than one that’s brilliant 70% of the time.

Agentic AI: Career-Defining Skill of 2026