The 2026 Agent (mental model)

CodeNow’s Anatomy view is built around Anthropic’s “2026 Agent” mental model — seven labeled regions every real agent has. This page is the printable cheat sheet.

The card


┌─────────────────────────────────────────────────────────┐
│  [icon] AgentName  workspace · projectName     ● 3 subagents
│  ┌─────┐                                                 
│  │skill│ ──────── chat conversation ─────                
│  ├─────┤  user: Schedule the Q3 review and prep brief   
│  │skill│  agent: I found three open slots — pick one... 
│  ├─────┤  ┌──────────────────────────────────────┐      
│  │skill│  │ Pick a slot          calendar mcp app │      ← rich widget
│  └─────┘  │ [Tue 14·10:00] [Wed 15·14:30] ...    │      
│           └──────────────────────────────────────┘      
│           ▶ composing notes =drive.search("Q2 review")  ← composes calls
│                                                          
│           Connectors                                     
│           [Cal] Google Calendar                          
│           [✉] Gmail                                     
│           [#] Slack                                      
│                                                          
│  ● Task · indexing repo · 2m  ● 14 servers connected  +  ← long-running + discover tools
└─────────────────────────────────────────────────────────┘

Seven regions

Region	Vocabulary	What it represents
Identity (top-left)	“icon · name · workspace”	Who this agent is
Subagents (top-right)	”● N subagents”	What it can delegate to
Skills rail (left)	“Applies skills · domain knowledge”	What it knows
Conversation (center)	“Rich interaction · MCP Apps”	What you said + what it said + interactive widgets
Composes calls (inline code)	“MCP + CLI as code”	What it’s doing right now (tool calls)
Connectors (bottom-center)	“Connected servers”	What it can reach
Long-running tasks (bottom-left status)	“async, surfaced”	What it’s working on in the background
Discovers tools (bottom-right `+`)	“progressive”	How it gets new capabilities

Why this view matters

Non-developers don’t have a mental picture of “an agent.” Saying “it’s a function that calls an LLM in a loop” doesn’t help. Showing them this card does.

The same card is also the demo screenshot for technical buyers. Anthropic’s research mockup showed this view as where AI app surfaces are heading; CodeNow built it as the actual UI of every agent.

Mapping CodeNow data → regions

CodeNow’s Anatomy panel populates each region from real project files:

Identity ← .codenow/agent-studio/agents/{slug}.json manifest
Subagents ← manifest.tools.length (clickable to drawer with usage stats)
Skills rail ← .codenow/skills.json, color-hashed deterministically per skill ID
Conversation ← last 10 recorder runs from .codenow/runs/{slug}/*.jsonl
Composes calls ← tool_use events from the latest run
Connectors ← per-project MCP from ~/.codenow/connections.json
Long-running tasks ← runs without a run.end event (still in flight)
Discovers tools ← + button opens the Visual Tool Builder

No mocks. Every region updates as the underlying file changes — no sync step.

Source

This mental model was distilled from Anthropic’s 2026 research mockup. The vocabulary maps cleanly onto every modern agent harness — Claude Agent SDK, Cursor SDK, Microsoft Agent Framework, Google ADK — so the same card explains agents built with any of them.