---
id: "action-simplify-agent-architecture"
type: "action-item"
source_timestamps: ["00:08:17", "00:08:45"]
tags: ["system-architecture", "debugging"]
related: ["claim-fancy-algorithms-fail-agents", "framework-rob-pike-agent-rules"]
speakers: ["Nate B. Jones"]
action: "Design simple, observable agent architectures rather than complex multi-agent routing graphs."
outcome: "Results in agentic systems that are significantly easier to debug and maintain."
sources: ["s41-nvidia-open-sourced"]
sourceVaultSlug: "s41-nvidia-open-sourced"
originDay: 41
---
# Simplify agent architectures to ensure debuggability

## Action

**Design simple, observable agent architectures rather than complex multi-agent routing graphs.**

## Why

Direct application of [[entity-rob-pike]]'s Rules 3 and 4 (see [[framework-rob-pike-agent-rules]] and [[claim-fancy-algorithms-fail-agents]]):
- Fancy algorithms are slow when N is small.
- Fancy algorithms are buggier than simple ones.

Most enterprise tasks have small N. Multi-agent routing graphs, deep prompt chains, and massive context stuffing become **black boxes** under failure — see [[quote-dont-get-fancy]].

## Concrete Steps

1. **Default to a single-agent loop** — react/observe/act, one model, one tool surface.
2. **Avoid orchestrator/worker patterns** unless you can show the simple version fails on a measured benchmark (see [[action-measure-before-optimizing]]).
3. **Make every step observable** — log the prompt, tool inputs, tool outputs, and model response for each turn.
4. **Constrain context aggressively** — short, structured context windows beat large unstructured ones.
5. **Promote complexity only when forced** — and only after baseline measurements prove the simple version is the bottleneck.

## When to Break the Rule

The enrichment overlay's counter-perspective is fair: at very large scale, multi-agent systems do outperform on tool-use benchmarks. Reach for hierarchical orchestration **only** when single-agent performance has demonstrably plateaued.

## Expected Outcome

- Failures become diagnosable in minutes instead of days.
- Maintenance burden drops sharply.
- Onboarding new engineers to the system becomes possible.

## See Also

- [[claim-fancy-algorithms-fail-agents]]
- [[framework-rob-pike-agent-rules]]
- [[quote-dont-get-fancy]]