The AI Landscape
2026.

A useful way to read the AI market in 2026 is along two axes — not one. Most maps you see stack everything on a single line of "more agentic" or "more capable." That hides where the real differences are.

The first axis is degree of delegation: a tool can answer, assist, execute a single task, or autonomously manage a whole process. The second is the environment of work: chat, documents, browser and computer, codebase, business systems, or the user's entire digital life.

Every product on the market sits somewhere on that grid. The interesting question for product people, founders and operators is not "which tool is best" but "which combination of delegation and environment matches the work I actually do."

AI is no longer one product category. It's a stack of layers, each with its own users, jobs, and risks.

Before the map, the words. Most confusion in AI conversations today comes from five terms used interchangeably when they describe different things.

Model. The brain. GPT, Claude, Gemini, Llama. A model can reason, generate and analyse — but it has no memory of you, no computer of its own, no access to your files or apps. The model is the engine. Everything else is the car around it.

Assistant. The interface to the brain. ChatGPT, Claude, Gemini, Microsoft Copilot. You ask, it answers. Search, memory, file uploads and image generation may be bolted on, but the fundamental unit of interaction is still request and response.

Copilot. Not an architecture — an interaction pattern. The human leads, the AI assists at each step: completes a line of code, rewrites a paragraph, explains an error, suggests the next move. Copilots multiply human throughput. They don't take the job.

Agent. You hand over an outcome, not a step. "Find the five best vendors, compare them, write a recommendation." The agent breaks the goal into steps, chooses tools, gathers information, takes actions, checks results, and loops until done.

Harness. The operating system around the agent. The harness decides what context the agent gets, which tools it can call, what it's allowed to change, where it runs, what it remembers, when it must ask permission, how the result is verified, what gets logged, and what happens when something fails.

The model gives intelligence. The agent gives an execution loop. The harness makes that loop safe, observable, and useful inside real work.

This last point is why the same underlying model can feel brilliant in one product and mediocre in another. The harness is doing most of the work you can see.

The poster above arranges the 2026 AI market into ten layers. Each layer answers a different question: who is it for, what work does it actually do, and how much of the loop does it own.

Layer

01

Foundation Models

The intelligence layer. Reasoning, language, code, multimodal understanding — the raw engines everything else is built on.

For who

Everyone, indirectly. Almost no one consumes raw models — they consume products built on top.

Examples

GPT-5 / 5.5, Claude 4.x, Gemini 3, Llama 4, Mistral Large.

PlayersOpenAIAnthropicGeminiMetaMistral

Layer

02

AI Assistants & Chat Interfaces

Conversational entry points to the models. You ask, they respond. Web search, file understanding, memory and image generation are bolted on.

For who

Everyone — personal and professional. The mainstream surface of AI.

Examples

ChatGPT, Claude, Gemini, Microsoft Copilot, Perplexity.

PlayersChatGPTClaudeGeminiCopilotPerplexity

Layer

03

Copilots

AI embedded inline in the tools you already use. Suggestions, autocompletes, in-context rewrites. The human stays in the driver's seat.

For who

Knowledge workers, developers, designers — anyone working inside an existing surface.

Examples

Microsoft 365 Copilot, Google Workspace Gemini, Notion AI, GitHub Copilot, Cursor (tab), Arc Max.

PlayersM365WorkspaceNotionCopilot

Layer

04

Task & Research Agents

Given a goal, they plan, browse, gather, execute, and produce a finished deliverable — not just an answer.

For who

Analysts, PMs, consultants, marketers, researchers.

Examples

ChatGPT Agent, Gemini Advanced Agent, Claude Projects, Manus, Glean, Perplexity Deep Research, Elicit, Consensus.

PlayersChatGPT AgentClaudeGeminiPerplexity

Layer

05

AI Coworkers (Desktop Agents)

Agents that operate across files, folders, and desktop applications like a junior employee — multi-step office work, no prompt-engineering required.

For who

Operations, finance, legal, marketing, founders — non-developer knowledge work.

Examples

Claude Cowork, Windows Copilot, Google Project Mariner, Rewind AI, plus verticals like Harvey (legal), Granola (meetings), Decagon (support).

PlayersClaudeWindowsGoogle

Layer

06

Developer Agents & AI-native IDEs

AI that reads the codebase, plans, edits across files, runs tests, debugs, and ships pull requests. Three flavours: inline copilots, agentic IDEs, terminal agents, and background cloud engineers.

For who

Engineers — from solo builders to platform teams.

Examples

Cursor, Windsurf, Replit Agent, JetBrains AI, Claude Code, Codex CLI, Gemini CLI, OpenCode, Pi, Devin, SWE-Agent, Metoro.

PlayersCopilotJetBrainsReplitClaude CodeCodex

Layer

07

Vibe / Prompt-to-App Builders

Natural language in, working web app out — frontend, backend, database, auth, hosting and deploy bundled together. The product (not the code) is the unit of work.

For who

Founders, PMs, designers, non-developers, anyone shipping prototypes or internal tools.

Examples

Lovable, Bolt.new, v0 by Vercel, Replit Agent, Framer AI.

Playersv0ReplitFramerBolt

Layer

08

Workflow & Business Automation Agents

AI agents living inside workflow engines and enterprise platforms. Move data between apps, classify, approve, notify, escalate.

For who

Operations, IT, RevOps, customer service, HR, finance — large teams and enterprises.

Examples

n8n, Zapier AI, Make, Microsoft Power Automate, Pipedream, ServiceNow AI, Salesforce Agentforce, Microsoft Copilot Studio, UiPath AI, Workday Illuminate.

Playersn8nZapierMakeSalesforceUiPath

Layer

09

Personal Agents (Persistent AI)

Always-on agents with long-term memory and cross-app reach. They know you, remember context, work on a schedule, and act proactively.

For who

Individuals, power users, early adopters.

Examples

OpenClaw, Poke, Rewind AI, Granola, Gemini Spark.

PlayersAppleGoogleOpenAI

Layer

10

Platforms, Frameworks & Harnesses

The substrate for building, deploying, observing and governing AI systems. SDKs, agent frameworks, evals, observability, delivery, and policy.

For who

AI engineers, platform teams, DevOps, MLOps, AI-native startups.

Examples

OpenAI Agents SDK, LangGraph, CrewAI, Anthropic Agent SDK, AutoGen, LangSmith, Arize Phoenix, Weights & Biases, Helicone, Traceloop, Harness, GitHub Actions, GitLab CI, Argo CD, Terraform.

PlayersLangChainAgents SDKAnthropicW&BTerraform

Layers 1–3 are about thinking with you. Layers 4–6 are about acting for you. Layers 7–9 are about building and running things on your behalf. Layer 10 is the substrate everyone else is quietly standing on.

Layer 6 deserves its own zoom because it's where most of the confusion in the market lives. "Coding agent" today means at least five different things.

Inline coding copilots. You write, the model suggests. GitHub Copilot autocomplete, JetBrains AI, Tabnine. Lowest autonomy, fastest adoption — boilerplate, tests, quick local fixes.

Agentic IDEs. The editor is built around the agent. Cursor, Windsurf, JetBrains AI Assistant, VS Code Insiders + Copilot, Replit Agent. You can hand off whole tasks, run several agents in parallel, and review diffs — without leaving the IDE.

Terminal / CLI agents. The shell is the interface. Claude Code, Codex CLI, Gemini CLI, OpenCode, Pi. They read the repo, run commands, edit files, run tests, and call MCP tools. Pi is the most explicit about being a minimal, programmable harness rather than an opinionated agent.

Cloud / background software engineers. Hand an agent an issue, it returns a pull request. Devin, GitHub Copilot Coding Agent, Codex cloud, Cursor background agents. Powerful — but only with good tests, clear acceptance criteria, and real review.

Vibe / prompt-to-app platforms. Layer 7 on the map, but worth contrasting. Lovable, Bolt, v0, Replit Agent. The unit of work isn't a file or a function — it's a product. UI, backend, database, auth and deploy come bundled.

Claude Code works with a software project. Lovable works with a product intention.

Once you can see the layers, picking the right one becomes a question, not a brand preference. The shortlist most teams need:

The map is not a leaderboard. Layer 2 isn't worse than Layer 10 — they're answers to different questions. The mistake most teams make is using one layer to do another layer's job, then concluding "AI doesn't work for us."

In 2024–2025 the dominant question was: who has the best model?

In 2026 the question has quietly changed to: which harness is this model running in?

Claude Code, Codex, OpenCode and Pi can run the same class of model — and produce very different results. The difference is in context management, tools, permissions, the execution loop, and how output gets verified. The harness is doing the work you mistake for model quality.

That's why Layer 10 — frameworks, harnesses, governance — is the layer that quietly decides whether an AI product is a demo or a system. It controls what the agent sees, what it can touch, what it remembers, what gets approved, what gets rolled back, and what gets audited.

The model is the engine. The harness is the car, the road, the seatbelt, and the brake.

For product teams building on top of AI, the practical implication is sharp: stop shopping for models, and start designing the harness around your specific work. That's where the durable product surface is.

If you collapse the whole map into a single chain, you get a stable way to talk about every new product launched next month:

Model → Assistant → Copilot → Task Agent → Coworker → Persistent Agent → Agent Platform → Harness & Governance.

Every tool is somewhere on that chain. Every job is best done by one or two adjacent links. Every confused conversation about "AI strategy" usually comes from arguing across links without knowing it.

The market will keep moving. New tools will arrive every week. But the chain — and the two-axis grid behind it — is stable enough to add new products to the map without redrawing it.

That, more than any specific model release, is the real shift of 2026.