Architecture chooser

Pick the pattern by ownership, uncertainty, and risk.

This is the reusable decision sheet for building Neurocode-like agents and production agent systems.

Single specialist

One agent owns a narrow task and directly uses its tools.

  1. Best for well-scoped expert jobs.
  2. Keep tool count small.
  3. Add guardrails around write actions.

Manager uses specialist as tool

The main agent keeps user-facing control and calls specialists as tools.

  1. Best when final answer needs one voice.
  2. Specialists return structured artifacts.
  3. Avoid recursive delegation.

Handoff

Ownership moves from one agent to another.

  1. Use when task responsibility changes.
  2. Pass reason and priority.
  3. Log handoff metadata.

Orchestrator-workers

Lead agent decomposes, workers explore independently, lead synthesizes.

  1. Best for research breadth.
  2. Give each worker objective and output schema.
  3. Merge evidence, not chat.

Evaluator-optimizer

Generator proposes, evaluator scores, optimizer revises.

  1. Best with a rubric or tests.
  2. Stop on threshold.
  3. Track cost per iteration.

Human approval gate

Agent proposes risky action; human approves before execution.

  1. Use for money, data deletion, security changes, external messages.
  2. Make approval diff clear.
  3. Record the decision.