The Full Agent Landscape in 2026

The agent ecosystem sorts cleanly into layers. Pick your layer and the choices get small; cross layers carelessly and you end up reinventing everything.

The big shifts since 2024

• Devin dropped from $500 to $20/month entry. Autonomous coding agents became consumer-priced.
• MCP became the de-facto standard with 1,200+ servers; backed by Anthropic, OpenAI, Google.
• Microsoft put AutoGen in maintenance mode in favor of its broader Microsoft Agent Framework.
• LangGraph passed CrewAI in GitHub stars — enterprise chose explicit state + audit trails.
• Anthropic acquired Vercept; Claude's OSWorld score jumped from <15% to 72.5%.
• OpenAI launched Workspace Agents as the replacement for Custom GPTs in orgs.
• UC Berkeley researchers published exploits for SWE-bench, WebArena, GAIA — every major benchmark can be gamed.

A decision matrix

Where the money is going

1. 1Enterprise workflow automation (Workspace Agents, LangGraph on private clouds).
2. 2Software engineering agents (Devin, Claude Code, Cursor Agents, Replit Agent).
3. 3Browser and computer-use agents for repetitive human work (Operator, Atlas).
4. 4Vertical agents (legal, medical, finance) fine-tuned on private corpora.
5. 5On-device agents on new NPUs (Apple Intelligence, Qualcomm Hexagon).

A quick mental map in code

// The four layers, from bottom to top:
type Stack = {
  // 1. Foundation model — the reasoning engine
  model: 'anthropic/claude-opus-4.7' | 'openai/gpt-5' | 'google/gemini-2.5' | 'local:qwen3.5:8b';

  // 2. Tool protocol — how the model reaches the world
  tools: 'mcp' | 'native-function-calling' | 'both';

  // 3. Framework — how you compose steps and state
  framework: 'langgraph' | 'crewai' | 'autogen' | 'openai-agents-sdk'
           | 'claude-code' | 'openclaw' | 'custom';

  // 4. Runtime — where it executes durably
  runtime: 'vercel-workflow' | 'temporal' | 'inngest' | 'self-hosted';
};

// Good modern stack: durable runtime + typed state + MCP tools + dotted model slug
const stack: Stack = {
  model: 'anthropic/claude-sonnet-4.6',
  tools: 'mcp',
  framework: 'langgraph',
  runtime: 'vercel-workflow',
};

The rest of this track goes deep. We start with the primitive — raw tool-use at the model API — then MCP, orchestration, and real production patterns.