Rise of the Agents: An AI Coding Ecosystem Map

Something shifted in 2025. AI coding tools stopped being autocomplete on steroids and started being actual agents — software that can take a task, break it down, write code, run tests, and ship a pull request while you sleep.

By early 2026, every major foundation lab has shipped a coding agent. Enterprises are building internal agent armies. And a new category of orchestration platforms has emerged to help teams manage the chaos.

Here's how I see the landscape.

Three Layers of the Stack

Infrastructure / Frameworks

At the bottom, two foundational categories:

Sandboxes like E2B, Daytona, and Modal provide isolated execution environments. Agents need somewhere safe to run code, and these platforms offer everything from Firecracker microVMs to GPU-accelerated containers.

Frameworks like LangChain, CrewAI, AutoGen, and Vercel AI SDK provide the building blocks. If you're building your own agents, you're probably starting here.

Agentic Engineering Platforms

The middle layer is where the real action is — three distinct approaches to the same problem:

Foundation Labs ship first-party agents with deep model integration. Claude from Anthropic, Codex from OpenAI, Gemini from Google. These have the tightest coupling with their respective models but lock you into a single provider.

Model-Agnostic Platforms like Tembo, Amp, Devin, and OpenCode sit above the foundation agents. They can orchestrate across providers, route work intelligently, and handle entire features end-to-end. This is where I'm personally spending time building — the platform layer is becoming critical as teams deploy multiple agents.

Enterprise In-House agents are the breakout story of 2026. Stripe's Minions ship 1,000+ PRs per week.^[1] Shopify, Spotify, StrongDM — every major tech company is building internal agent teams. These aren't just applications; they're becoming core engineering infrastructure.

Application

The top layer is where most users interact with agents:

Desktop Apps bring agents to individual developers. Cursor has 50%+ Fortune 500 adoption.^[2] Air from JetBrains, Codex App — the Mac menu bar is getting crowded.

Personal Agents like OpenClaw, Lindy, and Viktor extend beyond coding into general automation — scheduling, email, workflow orchestration.

Autonomous Coding Engines like Gastown, Ralph, and Pythagora push toward full autonomy with experimental approaches to end-to-end code generation.

Coding Agents like Aider, Cline, and Continue integrate into existing workflows — VS Code extensions, CLI tools, IDE plugins.

PM Agents like ChatPRD, Modem, Korey, and ProductForge focus on the upstream work — turning conversations into specs, managing roadmaps, coordinating between humans and agents.

What's Next

A few predictions:

Orchestration becomes table stakes. No serious team will run a single agent by 2027. The question is whether you build your own orchestration or buy it.

In-house agents go mainstream. What Stripe did, every company with 100+ engineers will attempt. The playbook is public now.

Desktop apps rise then consolidate. There are too many Mac coding apps. Expect acquisitions and shutdowns.

Foundation labs keep shipping. Claude, Codex, and Gemini will get dramatically better. The question is whether third-party tools can stay differentiated.

The rise of the agents is just beginning. The next 18 months will reshape how software gets built.

---

For deep dives on any of these tools, see my research section.