TheOnes.io

The One-Central-Brain Architecture: Why 20 Agent Tools Fragment Your Mind

2 min readWojciech Wiesner

  • ai-agents
  • orchestration
  • productivity
  • hermes
  • adhd

A fellow builder recently shared a metric tracker measuring activity across 20+ different agent CLIs—Cursor, Claude Code, Cline, Roo, Pi, Kimi, Codex, and more.

When I ran the tool on my own workstation over the last 90 days, the result was striking:

  • 2,223 local commits
  • 16,731 agent sessions
  • 24,694 prompts

But unlike most developers whose activity is split across a dozen disconnected apps, ~90% of my work lives inside a single central agent harness (Hermes Agent), backed by targeted specialist workers.

That contrast reveals one of the most critical debates in modern software engineering: Tool Fragmentation vs. Compound Context.

Trying every new agent CLI feels productive. You see a cool demo on Twitter, install a new CLI, play with it for two days, and feel like you are riding the bleeding edge.

In reality, tool hopping creates catastrophic context fragmentation:

  • Every new tool starts with a blank memory slate.
  • It doesn’t know your production infrastructure, your server layout, or your past deployment pitfalls.
  • It doesn’t know your coding style, your personal quirks, or the hard architectural decisions you made three months ago.
  • You spend your mental energy re-explaining the basics instead of solving the hard domain problem.

For someone with an AuDHD brain like mine, context switching is the ultimate productivity killer. Bouncing across five interfaces doesn’t double your output; it quarters your focus.

The Central Brain Architecture (Hub & Spoke)

Instead of fragmenting our workflow, we built an architecture centered around a single Stateful Cognitive Hub:

                  ┌───────────────────────────────┐
                  │    CENTRAL BRAIN (HERMES)     │
                  │ - Persistent Graph Memory     │
                  │ - Infrastructure Knowledge    │
                  │ - Unified Terminal & MCP Hub  │
                  └──────────────┬────────────────┘
                                 │
         ┌───────────────────────┼───────────────────────┐
         ▼                       ▼                       ▼
┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│ Specialized QA  │     │ Claude Code /   │     │ Background Task │
│  & Test Workers │     │  Opus Workers   │     │  & Cron Daemons │
└─────────────────┘     └─────────────────┘     └─────────────────┘

In this model, the Central Agent is your long-term partner:

  1. Durable Knowledge Compounds Over Time: It remembers that our Postgres database uses self-hosted Docker on Tailscale, that our deployment scripts require specific environment checks, and that destructive operations always require human sign-off.
  2. Dynamic Delegation Without Context Loss: When we need deep architectural reasoning or heavy refactoring, the central brain spawns subagents, passes them an exact, isolated brief, reviews their diffs, and integrates the results.
  3. Zero Mental Friction: Whether checking production logs, drafting technical RFCs, or deploying bugfixes, I talk to one interface that understands the entire map of my world.

The Verdict: Depth Beats Variety

The future of software engineering isn’t about collecting twenty different AI toys.

It’s about building a deeply integrated, governed system that preserves your focus, holds your operational memory, and multiplies your intent. When your agent remembers yesterday, today’s work starts ten steps ahead.