DESIGN TRADE-OFFS · SIDE BY SIDE

A 30-line agent loop is not wrong—it answers a smaller question

A tiny loop is ideal for learning tool calls. A harness addresses failure, recovery, governance, replacement, and long-term maintenance. Complexity moves into contracted boundaries.

Teaching-sized loop

while (true) {
  const reply = await model(messages, tools)
  messages.push(reply)
  if (!reply.toolCalls) break
  for (const call of reply.toolCalls) {
    messages.push(await tools[call.name](call.args))
  }
}

Local causality is clear. Add persistence, approval, cancellation, recovery, and surfaces, and every responsibility crowds the loop.

DeepSeek Harness

profile / bundle / patch
        ↓ compose
Context + scoped effects
        ↓ drive
Inbox → Turn → Step
        ↓ publish
SessionEvent + agent/*
        ↓ project
Web · CLI · ACP · SDK

The value is not fewer lines, but ownership, lifecycle, and test boundaries for each kind of change.

ConcernMinimal agentDeepSeek Harness

State truthIn-memory messages arrayAppend-only SessionEvent + projection

Execution unitOne loop iterationInbox → Turn → 1..n steps → settlement

Tool resultPush text directlyGuard pipeline + paired debt + settlement

PermissionIf inside each toolPreset → approval → monotonic guard

Long contextDrop old messagesCompaction + surface + spill + pair balance

Replace implementationEdit call sitesService definition → provider → consumer

Background workBuild another queueSubagent / Job / Workflow / Schedule contracts

Product compositionCopy config or forkProfile + Bundle + Patch + scoped disposal

USE THE SMALLEST SYSTEM THAT ANSWERS YOUR QUESTION

When should you graduate to a harness?

When you need cross-process recovery, multiple clients, auditable approval, replaceable capabilities, or parallel maintenance, explicit boundaries matter more than local line count.

Start from first principles