The shape of the stack — and who wins
The five-layer map has one more thing to teach you, and it's the part that matters if you ever build a company instead of a feature.
The layer count is contested. The shape isn't.
This course's five layers are one cut. Other writers slice the same territory into seven, or four, or nine. Don't memorize a number — the count is just an argument. What every version agrees on is the shape:
- The bottom — raw model intelligence, protocols, generic plumbing — commoditizes. It standardizes fast, and last year's hand-built version is next year's free library. A moat here is hard to dig.
- The top and middle — governance, the ecosystem of tools an agent can call, the economics of a task — is where durable advantage collects, because those take long to build and earn slow, real trust.
For a builder or PM the strategic question is never "which layer is the model." It is: given the model under me is a commodity, which layer am I actually defensible at?
Why most agents never ship
Industry surveys through 2025 and 2026 keep finding the same thing: only a small fraction of organizations have agents running at real scale. The gap is rarely the model — Intelligence is the layer that already works. The gap is the four layers around the loop. Most teams build a strong Action layer, demo it, and then spend the next two quarters building Governance and Orchestration before it is allowed near a customer.
When you scope your own agent in the later CLI-agent build, spend as much time on what the agent must not do as on what it can.
Read wider
This lesson compressed a fast-moving conversation. Primary sources, grouped by what they argue:
- The economy thesis — investor and vendor essays on the agentic AI economy, useful as arguments about where value accrues rather than as settled taxonomies.
- The annual industry survey — Nathan Benaich, State of AI Report 2025 (stateof.ai, published October 9, 2025) — the canonical year-in-review; the next edition ships October 2026.
- Primary incident reports — The Register / Tom's Hardware coverage of the April 27, 2026 PocketOS deletion by Cursor + Claude Opus 4.6; Engadget's reporting on the December 2025 AWS Kiro-coded environment-deletion outage; Jason Lemkin's X thread for the July 2025 Replit production-database wipe; AWS's October 20, 2025 us-east-1 post-incident summary. The first two are the documented version of the Governance / Orchestration failures cited in this chapter.
- Vendor docs as primary sources — Anthropic's "Building effective agents" post, OpenAI's tool-use docs, the MCP spec. The loop you actually have to build against, in the words of the people building the model.
- Skepticism layer — read the deployment-rate reporting from Gartner, Stanford HAI's AI Index, and McKinsey's State of AI surveys side by side; absolute numbers diverge but the trend (most orgs piloting, few in production) holds across methodologies.
The agentic-AI economy is being narrated in real time by people with money in it — investors, vendors, the labs themselves. Read every layer diagram, the one in this lesson included, as an argument. The loop you wrote in this chapter is the fact. The map is just someone's opinion about where the loop is standing.
The shape of the stack — and who wins
The five-layer map has one more thing to teach you, and it's the part that matters if you ever build a company instead of a feature.
The layer count is contested. The shape isn't.
This course's five layers are one cut. Other writers slice the same territory into seven, or four, or nine. Don't memorize a number — the count is just an argument. What every version agrees on is the shape:
- The bottom — raw model intelligence, protocols, generic plumbing — commoditizes. It standardizes fast, and last year's hand-built version is next year's free library. A moat here is hard to dig.
- The top and middle — governance, the ecosystem of tools an agent can call, the economics of a task — is where durable advantage collects, because those take long to build and earn slow, real trust.
For a builder or PM the strategic question is never "which layer is the model." It is: given the model under me is a commodity, which layer am I actually defensible at?
Why most agents never ship
Industry surveys through 2025 and 2026 keep finding the same thing: only a small fraction of organizations have agents running at real scale. The gap is rarely the model — Intelligence is the layer that already works. The gap is the four layers around the loop. Most teams build a strong Action layer, demo it, and then spend the next two quarters building Governance and Orchestration before it is allowed near a customer.
When you scope your own agent in the later CLI-agent build, spend as much time on what the agent must not do as on what it can.
Read wider
This lesson compressed a fast-moving conversation. Primary sources, grouped by what they argue:
- The economy thesis — investor and vendor essays on the agentic AI economy, useful as arguments about where value accrues rather than as settled taxonomies.
- The annual industry survey — Nathan Benaich, State of AI Report 2025 (stateof.ai, published October 9, 2025) — the canonical year-in-review; the next edition ships October 2026.
- Primary incident reports — The Register / Tom's Hardware coverage of the April 27, 2026 PocketOS deletion by Cursor + Claude Opus 4.6; Engadget's reporting on the December 2025 AWS Kiro-coded environment-deletion outage; Jason Lemkin's X thread for the July 2025 Replit production-database wipe; AWS's October 20, 2025 us-east-1 post-incident summary. The first two are the documented version of the Governance / Orchestration failures cited in this chapter.
- Vendor docs as primary sources — Anthropic's "Building effective agents" post, OpenAI's tool-use docs, the MCP spec. The loop you actually have to build against, in the words of the people building the model.
- Skepticism layer — read the deployment-rate reporting from Gartner, Stanford HAI's AI Index, and McKinsey's State of AI surveys side by side; absolute numbers diverge but the trend (most orgs piloting, few in production) holds across methodologies.
The agentic-AI economy is being narrated in real time by people with money in it — investors, vendors, the labs themselves. Read every layer diagram, the one in this lesson included, as an argument. The loop you wrote in this chapter is the fact. The map is just someone's opinion about where the loop is standing.
The shape of the stack — and who wins
The five-layer map has one more thing to teach you, and it's the part that matters if you ever build a company instead of a feature.
The layer count is contested. The shape isn't.
This course's five layers are one cut. Other writers slice the same territory into seven, or four, or nine. Don't memorize a number — the count is just an argument. What every version agrees on is the shape:
- The bottom — raw model intelligence, protocols, generic plumbing — commoditizes. It standardizes fast, and last year's hand-built version is next year's free library. A moat here is hard to dig.
- The top and middle — governance, the ecosystem of tools an agent can call, the economics of a task — is where durable advantage collects, because those take long to build and earn slow, real trust.
For a builder or PM the strategic question is never "which layer is the model." It is: given the model under me is a commodity, which layer am I actually defensible at?
Why most agents never ship
Industry surveys through 2025 and 2026 keep finding the same thing: only a small fraction of organizations have agents running at real scale. The gap is rarely the model — Intelligence is the layer that already works. The gap is the four layers around the loop. Most teams build a strong Action layer, demo it, and then spend the next two quarters building Governance and Orchestration before it is allowed near a customer.
When you scope your own agent , spend as much time on what the agent must not do as on what it can.
Read wider
This lesson compressed a fast-moving conversation. Primary sources, grouped by what they argue:
- The economy thesis — investor and vendor essays on the agentic AI economy, useful as arguments about where value accrues rather than as settled taxonomies.
- The annual industry survey — Nathan Benaich, State of AI Report 2025 (stateof.ai, published October 9, 2025) — the canonical year-in-review; the next edition ships October 2026.
- Primary incident reports — The Register / Tom's Hardware coverage of the April 27, 2026 PocketOS deletion by Cursor + Claude Opus 4.6; Engadget's reporting on the December 2025 AWS Kiro-coded environment-deletion outage; Jason Lemkin's X thread for the July 2025 Replit production-database wipe; AWS's October 20, 2025 us-east-1 post-incident summary. The first two are the documented version of the Governance / Orchestration failures cited in this chapter.
- Vendor docs as primary sources — Anthropic's "Building effective agents" post, OpenAI's tool-use docs, the MCP spec. The loop you actually have to build against, in the words of the people building the model.
- Skepticism layer — read the deployment-rate reporting from Gartner, Stanford HAI's AI Index, and McKinsey's State of AI surveys side by side; absolute numbers diverge but the trend (most orgs piloting, few in production) holds across methodologies.
The agentic-AI economy is being narrated in real time by people with money in it — investors, vendors, the labs themselves. Read every layer diagram, the one in this lesson included, as an argument. The loop you wrote in this chapter is the fact. The map is just someone's opinion about where the loop is standing.
The shape of the stack — and who wins
The five-layer map has one more thing to teach you, and it's the part that matters if you ever build a company instead of a feature.
The layer count is contested. The shape isn't.
This course's five layers are one cut. Other writers slice the same territory into seven, or four, or nine. Don't memorize a number — the count is just an argument. What every version agrees on is the shape:
- The bottom — raw model intelligence, protocols, generic plumbing — commoditizes. It standardizes fast, and last year's hand-built version is next year's free library. A moat here is hard to dig.
- The top and middle — governance, the ecosystem of tools an agent can call, the economics of a task — is where durable advantage collects, because those take long to build and earn slow, real trust.
For a builder or PM the strategic question is never "which layer is the model." It is: given the model under me is a commodity, which layer am I actually defensible at?
Why most agents never ship
Industry surveys through 2025 and 2026 keep finding the same thing: only a small fraction of organizations have agents running at real scale. The gap is rarely the model — Intelligence is the layer that already works. The gap is the four layers around the loop. Most teams build a strong Action layer, demo it, and then spend the next two quarters building Governance and Orchestration before it is allowed near a customer.
When you scope your own agent , spend as much time on what the agent must not do as on what it can.
Read wider
This lesson compressed a fast-moving conversation. Primary sources, grouped by what they argue:
- The economy thesis — investor and vendor essays on the agentic AI economy, useful as arguments about where value accrues rather than as settled taxonomies.
- The annual industry survey — Nathan Benaich, State of AI Report 2025 (stateof.ai, published October 9, 2025) — the canonical year-in-review; the next edition ships October 2026.
- Primary incident reports — The Register / Tom's Hardware coverage of the April 27, 2026 PocketOS deletion by Cursor + Claude Opus 4.6; Engadget's reporting on the December 2025 AWS Kiro-coded environment-deletion outage; Jason Lemkin's X thread for the July 2025 Replit production-database wipe; AWS's October 20, 2025 us-east-1 post-incident summary. The first two are the documented version of the Governance / Orchestration failures cited in this chapter.
- Vendor docs as primary sources — Anthropic's "Building effective agents" post, OpenAI's tool-use docs, the MCP spec. The loop you actually have to build against, in the words of the people building the model.
- Skepticism layer — read the deployment-rate reporting from Gartner, Stanford HAI's AI Index, and McKinsey's State of AI surveys side by side; absolute numbers diverge but the trend (most orgs piloting, few in production) holds across methodologies.
The agentic-AI economy is being narrated in real time by people with money in it — investors, vendors, the labs themselves. Read every layer diagram, the one in this lesson included, as an argument. The loop you wrote in this chapter is the fact. The map is just someone's opinion about where the loop is standing.
The shape of the stack — and who wins
The five-layer map has one more thing to teach you, and it's the part that matters if you ever build a company instead of a feature.
The layer count is contested. The shape isn't.
This course's five layers are one cut. Other writers slice the same territory into seven, or four, or nine. Don't memorize a number — the count is just an argument. What every version agrees on is the shape:
- The bottom — raw model intelligence, protocols, generic plumbing — commoditizes. It standardizes fast, and last year's hand-built version is next year's free library. A moat here is hard to dig.
- The top and middle — governance, the ecosystem of tools an agent can call, the economics of a task — is where durable advantage collects, because those take long to build and earn slow, real trust.
For a builder or PM the strategic question is never "which layer is the model." It is: given the model under me is a commodity, which layer am I actually defensible at?
Why most agents never ship
Industry surveys through 2025 and 2026 keep finding the same thing: only a small fraction of organizations have agents running at real scale. The gap is rarely the model — Intelligence is the layer that already works. The gap is the four layers around the loop. Most teams build a strong Action layer, demo it, and then spend the next two quarters building Governance and Orchestration before it is allowed near a customer.
When you scope your own agent , spend as much time on what the agent must not do as on what it can.
Read wider
This lesson compressed a fast-moving conversation. Primary sources, grouped by what they argue:
- The economy thesis — investor and vendor essays on the agentic AI economy, useful as arguments about where value accrues rather than as settled taxonomies.
- The annual industry survey — Nathan Benaich, State of AI Report 2025 (stateof.ai, published October 9, 2025) — the canonical year-in-review; the next edition ships October 2026.
- Primary incident reports — The Register / Tom's Hardware coverage of the April 27, 2026 PocketOS deletion by Cursor + Claude Opus 4.6; Engadget's reporting on the December 2025 AWS Kiro-coded environment-deletion outage; Jason Lemkin's X thread for the July 2025 Replit production-database wipe; AWS's October 20, 2025 us-east-1 post-incident summary. The first two are the documented version of the Governance / Orchestration failures cited in this chapter.
- Vendor docs as primary sources — Anthropic's "Building effective agents" post, OpenAI's tool-use docs, the MCP spec. The loop you actually have to build against, in the words of the people building the model.
- Skepticism layer — read the deployment-rate reporting from Gartner, Stanford HAI's AI Index, and McKinsey's State of AI surveys side by side; absolute numbers diverge but the trend (most orgs piloting, few in production) holds across methodologies.
The agentic-AI economy is being narrated in real time by people with money in it — investors, vendors, the labs themselves. Read every layer diagram, the one in this lesson included, as an argument. The loop you wrote in this chapter is the fact. The map is just someone's opinion about where the loop is standing.
The shape of the stack — and who wins
The five-layer map has one more thing to teach you, and it's the part that matters if you ever build a company instead of a feature.
The layer count is contested. The shape isn't.
This course's five layers are one cut. Other writers slice the same territory into seven, or four, or nine. Don't memorize a number — the count is just an argument. What every version agrees on is the shape:
- The bottom — raw model intelligence, protocols, generic plumbing — commoditizes. It standardizes fast, and last year's hand-built version is next year's free library. A moat here is hard to dig.
- The top and middle — governance, the ecosystem of tools an agent can call, the economics of a task — is where durable advantage collects, because those take long to build and earn slow, real trust.
For a builder or PM the strategic question is never "which layer is the model." It is: given the model under me is a commodity, which layer am I actually defensible at?
Why most agents never ship
Industry surveys through 2025 and 2026 keep finding the same thing: only a small fraction of organizations have agents running at real scale. The gap is rarely the model — Intelligence is the layer that already works. The gap is the four layers around the loop. Most teams build a strong Action layer, demo it, and then spend the next two quarters building Governance and Orchestration before it is allowed near a customer.
When you scope your own agent , spend as much time on what the agent must not do as on what it can.
Read wider
This lesson compressed a fast-moving conversation. Primary sources, grouped by what they argue:
- The economy thesis — investor and vendor essays on the agentic AI economy, useful as arguments about where value accrues rather than as settled taxonomies.
- The annual industry survey — Nathan Benaich, State of AI Report 2025 (stateof.ai, published October 9, 2025) — the canonical year-in-review; the next edition ships October 2026.
- Primary incident reports — The Register / Tom's Hardware coverage of the April 27, 2026 PocketOS deletion by Cursor + Claude Opus 4.6; Engadget's reporting on the December 2025 AWS Kiro-coded environment-deletion outage; Jason Lemkin's X thread for the July 2025 Replit production-database wipe; AWS's October 20, 2025 us-east-1 post-incident summary. The first two are the documented version of the Governance / Orchestration failures cited in this chapter.
- Vendor docs as primary sources — Anthropic's "Building effective agents" post, OpenAI's tool-use docs, the MCP spec. The loop you actually have to build against, in the words of the people building the model.
- Skepticism layer — read the deployment-rate reporting from Gartner, Stanford HAI's AI Index, and McKinsey's State of AI surveys side by side; absolute numbers diverge but the trend (most orgs piloting, few in production) holds across methodologies.
The agentic-AI economy is being narrated in real time by people with money in it — investors, vendors, the labs themselves. Read every layer diagram, the one in this lesson included, as an argument. The loop you wrote in this chapter is the fact. The map is just someone's opinion about where the loop is standing.
The shape of the stack — and who wins
The five-layer map has one more thing to teach you, and it's the part that matters if you ever build a company instead of a feature.
The layer count is contested. The shape isn't.
This course's five layers are one cut. Other writers slice the same territory into seven, or four, or nine. Don't memorize a number — the count is just an argument. What every version agrees on is the shape:
- The bottom — raw model intelligence, protocols, generic plumbing — commoditizes. It standardizes fast, and last year's hand-built version is next year's free library. A moat here is hard to dig.
- The top and middle — governance, the ecosystem of tools an agent can call, the economics of a task — is where durable advantage collects, because those take long to build and earn slow, real trust.
For a builder or PM the strategic question is never "which layer is the model." It is: given the model under me is a commodity, which layer am I actually defensible at?
Why most agents never ship
Industry surveys through 2025 and 2026 keep finding the same thing: only a small fraction of organizations have agents running at real scale. The gap is rarely the model — Intelligence is the layer that already works. The gap is the four layers around the loop. Most teams build a strong Action layer, demo it, and then spend the next two quarters building Governance and Orchestration before it is allowed near a customer.
When you scope your own agent , spend as much time on what the agent must not do as on what it can.
Read wider
This lesson compressed a fast-moving conversation. Primary sources, grouped by what they argue:
- The economy thesis — investor and vendor essays on the agentic AI economy, useful as arguments about where value accrues rather than as settled taxonomies.
- The annual industry survey — Nathan Benaich, State of AI Report 2025 (stateof.ai, published October 9, 2025) — the canonical year-in-review; the next edition ships October 2026.
- Primary incident reports — The Register / Tom's Hardware coverage of the April 27, 2026 PocketOS deletion by Cursor + Claude Opus 4.6; Engadget's reporting on the December 2025 AWS Kiro-coded environment-deletion outage; Jason Lemkin's X thread for the July 2025 Replit production-database wipe; AWS's October 20, 2025 us-east-1 post-incident summary. The first two are the documented version of the Governance / Orchestration failures cited in this chapter.
- Vendor docs as primary sources — Anthropic's "Building effective agents" post, OpenAI's tool-use docs, the MCP spec. The loop you actually have to build against, in the words of the people building the model.
- Skepticism layer — read the deployment-rate reporting from Gartner, Stanford HAI's AI Index, and McKinsey's State of AI surveys side by side; absolute numbers diverge but the trend (most orgs piloting, few in production) holds across methodologies.
The agentic-AI economy is being narrated in real time by people with money in it — investors, vendors, the labs themselves. Read every layer diagram, the one in this lesson included, as an argument. The loop you wrote in this chapter is the fact. The map is just someone's opinion about where the loop is standing.
The shape of the stack — and who wins
The five-layer map has one more thing to teach you, and it's the part that matters if you ever build a company instead of a feature.
The layer count is contested. The shape isn't.
This course's five layers are one cut. Other writers slice the same territory into seven, or four, or nine. Don't memorize a number — the count is just an argument. What every version agrees on is the shape:
- The bottom — raw model intelligence, protocols, generic plumbing — commoditizes. It standardizes fast, and last year's hand-built version is next year's free library. A moat here is hard to dig.
- The top and middle — governance, the ecosystem of tools an agent can call, the economics of a task — is where durable advantage collects, because those take long to build and earn slow, real trust.
For a builder or PM the strategic question is never "which layer is the model." It is: given the model under me is a commodity, which layer am I actually defensible at?
Why most agents never ship
Industry surveys through 2025 and 2026 keep finding the same thing: only a small fraction of organizations have agents running at real scale. The gap is rarely the model — Intelligence is the layer that already works. The gap is the four layers around the loop. Most teams build a strong Action layer, demo it, and then spend the next two quarters building Governance and Orchestration before it is allowed near a customer.
When you scope your own agent , spend as much time on what the agent must not do as on what it can.
Read wider
This lesson compressed a fast-moving conversation. Primary sources, grouped by what they argue:
- The economy thesis — investor and vendor essays on the agentic AI economy, useful as arguments about where value accrues rather than as settled taxonomies.
- The annual industry survey — Nathan Benaich, State of AI Report 2025 (stateof.ai, published October 9, 2025) — the canonical year-in-review; the next edition ships October 2026.
- Primary incident reports — The Register / Tom's Hardware coverage of the April 27, 2026 PocketOS deletion by Cursor + Claude Opus 4.6; Engadget's reporting on the December 2025 AWS Kiro-coded environment-deletion outage; Jason Lemkin's X thread for the July 2025 Replit production-database wipe; AWS's October 20, 2025 us-east-1 post-incident summary. The first two are the documented version of the Governance / Orchestration failures cited in this chapter.
- Vendor docs as primary sources — Anthropic's "Building effective agents" post, OpenAI's tool-use docs, the MCP spec. The loop you actually have to build against, in the words of the people building the model.
- Skepticism layer — read the deployment-rate reporting from Gartner, Stanford HAI's AI Index, and McKinsey's State of AI surveys side by side; absolute numbers diverge but the trend (most orgs piloting, few in production) holds across methodologies.
The agentic-AI economy is being narrated in real time by people with money in it — investors, vendors, the labs themselves. Read every layer diagram, the one in this lesson included, as an argument. The loop you wrote in this chapter is the fact. The map is just someone's opinion about where the loop is standing.
The shape of the stack — and who wins
The five-layer map has one more thing to teach you, and it's the part that matters if you ever build a company instead of a feature.
The layer count is contested. The shape isn't.
This course's five layers are one cut. Other writers slice the same territory into seven, or four, or nine. Don't memorize a number — the count is just an argument. What every version agrees on is the shape:
- The bottom — raw model intelligence, protocols, generic plumbing — commoditizes. It standardizes fast, and last year's hand-built version is next year's free library. A moat here is hard to dig.
- The top and middle — governance, the ecosystem of tools an agent can call, the economics of a task — is where durable advantage collects, because those take long to build and earn slow, real trust.
For a builder or PM the strategic question is never "which layer is the model." It is: given the model under me is a commodity, which layer am I actually defensible at?
Why most agents never ship
Industry surveys through 2025 and 2026 keep finding the same thing: only a small fraction of organizations have agents running at real scale. The gap is rarely the model — Intelligence is the layer that already works. The gap is the four layers around the loop. Most teams build a strong Action layer, demo it, and then spend the next two quarters building Governance and Orchestration before it is allowed near a customer.
When you scope your own agent , spend as much time on what the agent must not do as on what it can.
Read wider
This lesson compressed a fast-moving conversation. Primary sources, grouped by what they argue:
- The economy thesis — investor and vendor essays on the agentic AI economy, useful as arguments about where value accrues rather than as settled taxonomies.
- The annual industry survey — Nathan Benaich, State of AI Report 2025 (stateof.ai, published October 9, 2025) — the canonical year-in-review; the next edition ships October 2026.
- Primary incident reports — The Register / Tom's Hardware coverage of the April 27, 2026 PocketOS deletion by Cursor + Claude Opus 4.6; Engadget's reporting on the December 2025 AWS Kiro-coded environment-deletion outage; Jason Lemkin's X thread for the July 2025 Replit production-database wipe; AWS's October 20, 2025 us-east-1 post-incident summary. The first two are the documented version of the Governance / Orchestration failures cited in this chapter.
- Vendor docs as primary sources — Anthropic's "Building effective agents" post, OpenAI's tool-use docs, the MCP spec. The loop you actually have to build against, in the words of the people building the model.
- Skepticism layer — read the deployment-rate reporting from Gartner, Stanford HAI's AI Index, and McKinsey's State of AI surveys side by side; absolute numbers diverge but the trend (most orgs piloting, few in production) holds across methodologies.
The agentic-AI economy is being narrated in real time by people with money in it — investors, vendors, the labs themselves. Read every layer diagram, the one in this lesson included, as an argument. The loop you wrote in this chapter is the fact. The map is just someone's opinion about where the loop is standing.
The shape of the stack — and who wins
The five-layer map has one more thing to teach you, and it's the part that matters if you ever build a company instead of a feature.
The layer count is contested. The shape isn't.
This course's five layers are one cut. Other writers slice the same territory into seven, or four, or nine. Don't memorize a number — the count is just an argument. What every version agrees on is the shape:
- The bottom — raw model intelligence, protocols, generic plumbing — commoditizes. It standardizes fast, and last year's hand-built version is next year's free library. A moat here is hard to dig.
- The top and middle — governance, the ecosystem of tools an agent can call, the economics of a task — is where durable advantage collects, because those take long to build and earn slow, real trust.
For a builder or PM the strategic question is never "which layer is the model." It is: given the model under me is a commodity, which layer am I actually defensible at?
Why most agents never ship
Industry surveys through 2025 and 2026 keep finding the same thing: only a small fraction of organizations have agents running at real scale. The gap is rarely the model — Intelligence is the layer that already works. The gap is the four layers around the loop. Most teams build a strong Action layer, demo it, and then spend the next two quarters building Governance and Orchestration before it is allowed near a customer.
When you scope your own agent , spend as much time on what the agent must not do as on what it can.
Read wider
This lesson compressed a fast-moving conversation. Primary sources, grouped by what they argue:
- The economy thesis — investor and vendor essays on the agentic AI economy, useful as arguments about where value accrues rather than as settled taxonomies.
- The annual industry survey — Nathan Benaich, State of AI Report 2025 (stateof.ai, published October 9, 2025) — the canonical year-in-review; the next edition ships October 2026.
- Primary incident reports — The Register / Tom's Hardware coverage of the April 27, 2026 PocketOS deletion by Cursor + Claude Opus 4.6; Engadget's reporting on the December 2025 AWS Kiro-coded environment-deletion outage; Jason Lemkin's X thread for the July 2025 Replit production-database wipe; AWS's October 20, 2025 us-east-1 post-incident summary. The first two are the documented version of the Governance / Orchestration failures cited in this chapter.
- Vendor docs as primary sources — Anthropic's "Building effective agents" post, OpenAI's tool-use docs, the MCP spec. The loop you actually have to build against, in the words of the people building the model.
- Skepticism layer — read the deployment-rate reporting from Gartner, Stanford HAI's AI Index, and McKinsey's State of AI surveys side by side; absolute numbers diverge but the trend (most orgs piloting, few in production) holds across methodologies.
The agentic-AI economy is being narrated in real time by people with money in it — investors, vendors, the labs themselves. Read every layer diagram, the one in this lesson included, as an argument. The loop you wrote in this chapter is the fact. The map is just someone's opinion about where the loop is standing.
The shape of the stack — and who wins
The five-layer map has one more thing to teach you, and it's the part that matters if you ever build a company instead of a feature.
The layer count is contested. The shape isn't.
This course's five layers are one cut. Other writers slice the same territory into seven, or four, or nine. Don't memorize a number — the count is just an argument. What every version agrees on is the shape:
- The bottom — raw model intelligence, protocols, generic plumbing — commoditizes. It standardizes fast, and last year's hand-built version is next year's free library. A moat here is hard to dig.
- The top and middle — governance, the ecosystem of tools an agent can call, the economics of a task — is where durable advantage collects, because those take long to build and earn slow, real trust.
For a builder or PM the strategic question is never "which layer is the model." It is: given the model under me is a commodity, which layer am I actually defensible at?
Why most agents never ship
Industry surveys through 2025 and 2026 keep finding the same thing: only a small fraction of organizations have agents running at real scale. The gap is rarely the model — Intelligence is the layer that already works. The gap is the four layers around the loop. Most teams build a strong Action layer, demo it, and then spend the next two quarters building Governance and Orchestration before it is allowed near a customer.
When you scope your own agent , spend as much time on what the agent must not do as on what it can.
Read wider
This lesson compressed a fast-moving conversation. Primary sources, grouped by what they argue:
- The economy thesis — investor and vendor essays on the agentic AI economy, useful as arguments about where value accrues rather than as settled taxonomies.
- The annual industry survey — Nathan Benaich, State of AI Report 2025 (stateof.ai, published October 9, 2025) — the canonical year-in-review; the next edition ships October 2026.
- Primary incident reports — The Register / Tom's Hardware coverage of the April 27, 2026 PocketOS deletion by Cursor + Claude Opus 4.6; Engadget's reporting on the December 2025 AWS Kiro-coded environment-deletion outage; Jason Lemkin's X thread for the July 2025 Replit production-database wipe; AWS's October 20, 2025 us-east-1 post-incident summary. The first two are the documented version of the Governance / Orchestration failures cited in this chapter.
- Vendor docs as primary sources — Anthropic's "Building effective agents" post, OpenAI's tool-use docs, the MCP spec. The loop you actually have to build against, in the words of the people building the model.
- Skepticism layer — read the deployment-rate reporting from Gartner, Stanford HAI's AI Index, and McKinsey's State of AI surveys side by side; absolute numbers diverge but the trend (most orgs piloting, few in production) holds across methodologies.
The agentic-AI economy is being narrated in real time by people with money in it — investors, vendors, the labs themselves. Read every layer diagram, the one in this lesson included, as an argument. The loop you wrote in this chapter is the fact. The map is just someone's opinion about where the loop is standing.