Why four agents beat one
Confirmation bias, context drift and the babysitting tax are structural failures of single-agent loops. Independence is the fix.
--- title: "Why four agents beat one" description: "Confirmation bias, context drift and the babysitting tax are structural failures of single-agent loops. Independence is the fix." pubDate: "2026-08-25" author: "Roshan Ravani" tags: ["architecture"] ---
Every single-agent coding tool — no matter how good the model — runs one long conversation. Three failures follow from that shape alone:
Confirmation bias. An agent that wrote the code reviews its own reasoning. It never truly challenges its assumptions because it is its own audience.
Context drift. Quality degrades as the conversation grows. Every earlier token competes with the task at hand.
The babysitting tax. You become the reviewer, steerer and verifier of a stream of edits — the opposite of delegation.
The artifact boundary
Niki splits the work into Planner, Coder, Tester and Reviewer — agents that share no conversation. They exchange typed artifacts: a TaskSpec, a unified diff, a test report, a verdict. Each artifact is JSON-schema validated before the next stage consumes it.
The result: a Tester that never saw the Coder's reasoning can only judge the change. A Reviewer that never saw the Planner's intent can only judge whether the spec was met. The biases have no channel to travel through.
Independence is testable
This isn't philosophy — it's enforced in code. Isolation records mirror the wiring, Red agents receive evidence-only projections, and shared-model review raises warnings. See the claims audit for the code-level evidence.