---
id: e3
slug: when-flat-beats-fancy-killing-agent-complexity
title: "When Flat Beats Fancy: Why We Killed Our Multi-Agent Hierarchy, Semantic DNS, and Lineage Commissions"
date: September 2026
tags: [Agent Architecture, Negative Results, Sovereign Firm, LLM Wiki]
canonical: https://amyote.com/#essay-when-flat-beats-fancy-killing-agent-complexity
---

# When Flat Beats Fancy: Why We Killed Our Multi-Agent Hierarchy, Semantic DNS, and Lineage Commissions

> Three negative results from our lab, and how Coase, Simon, and a shared Markdown Wiki explain why most multi-agent org charts are pure overhead.

**Date:** September 2026 · **Reading time:** 4 min read  
**Tags:** Agent Architecture, Negative Results, Sovereign Firm, LLM Wiki  

Walk into any enterprise AI pitch in 2026 and you will meet the same diagram: a Supervisor Agent delegating to a Curator Agent, queries routed through a semantic tree, sub-agents paid multi-level commissions. It looks wonderfully busy. We built all three in code, which is how we know that our own benchmarks then made us kill or flatten every one. [1]

The lab rule is Measured Architecture Complexity: no new agent role, memory layer or routing tree survives unless it beats a dead-simple baseline at equal token cost and latency. Most org charts, it turns out, do not. Ours certainly didn’t.

> **Pivot Question:** What happens when you pit a multi-layered agent organization against a flat worker pool writing directly to a source-backed Markdown Wiki?

In AgentBoard (our offline replay of real Home PiPi assistant events plus a synthetic agency benchmark), we tested whether a shared message Board and a dedicated Curator agent improved collective memory over plain Wiki files. Result: both achieved 916/916 on reader comprehension and 749/749 on organization tasks, but the flat B′ architecture (a single CEO spawning short-lived sub-agents that write source-backed fact lines directly to the Wiki, with no LLM rewrite pass and no Curator) was 16% cheaper. We deleted the Curator. [2]

In DNS_FOX, we tested whether a hierarchical Semantic DNS tree could route agent memory queries better than flat vector search across 827 memories, 68 resources, 3 embedding models, and 9 tree topologies. Verdict: KILL. The semantic tree lost to flat search on every single routing metric while adding maintenance fragility.

In AI MLM (Lineage Incentives Research), a pre-registered 40-seed MARL study compared flat TEAM rewards, DIRECT bounties, and multi-level LINEAGE commissions for agents that spawn other agents. Final v2 result: lineage commissions showed no significant advantage over direct rewards and introduced the pathology of irreversible spawning collapse. Why do all three results point the same way? As Herbert Simon noted, and as recent work on the Agentic & Sovereign Firm highlights, centralizing a verifiable System of Record (like our 370-page LLM Wiki) decentralizes execution. When every agent reads and writes provenance-checked facts to a shared ledger, middle-manager agents become pure tax. [3]

## How to build agent systems that survive contact with production

1. Replace middle-manager agents with a System of Record. Instead of chaining three LLMs to summarize each other’s chat logs, enforce a deterministic write contract: every fact needs a source ID, and newer facts supersede older ones in an append-only log.
2. Gate every experiment on a competence baseline. In AI MLM, our complex expansion environment (E1u) failed a simple script-baseline competence check, saving us from drawing false conclusions about incentive math on a broken task.
3. Reserve hierarchy for conflict resolution at scale. Extra ontology layers (like Databricks OntoRank) only earn their keep when resolving conflicting definitions across massive graphs (>65k tokens), not in everyday team workflows.

## Margin Notes & Caveats

- **[1] Why Publish Negative Results?:** 67% of corporate AI pilots stall in "Pilot Purgatory" because teams mistake architectural ceremony for capability. Publishing clean negative results saves months of engineering dead ends.
- **[2] Boundary of the AgentBoard Result:** The 916/916 and 749/749 parity holds for structured fact lines with an explicit source-reliability rule under ~65k tokens. It does not claim that raw, uncurated free-text chat logs scale indefinitely.
- **[3] Sovereign Memory (Gray Fox & amyote OS):** We apply the same principle in amyote OS (where a Python CLI validates bet logs with zero LLM calls inside the loop) and Gray Fox (where canonical state lives on Ethereum while the LLM Cortex is disposable).

## Appendix: Empirical Evidence

### Did complex agent routing, curation, or lineage incentives beat flat baselines?

**Takeaway:** Across all three completed internal experiments, flat architectures with explicit provenance rules matched or beat hierarchical alternatives at lower cost and lower failure risk.

*Summary of locked internal evaluation runs (Sept–Oct 2026). Baselines marked with *.*

| Experiment | Complex Mechanism Tested | Flat Baseline* | Measured Result | Engineering Decision |
| --- | --- | --- | --- | --- |
| AgentBoard (Home PiPi Replay) | Board + Curator Role + LLM Rewrite (Org A) | Direct fact lines + Reliability Rule (Org B′)* | 916/916 reader; 749/749 org parity; B′ is ~16% cheaper | KILL Curator; write source-backed facts to Wiki |
| DNS_FOX (Semantic Routing) | Hierarchical Semantic DNS Tree (9 shapes × 3 models) | Flat Vector Search + Explicit Policy* | Tree lost to flat search across all 105 queries / 827 memories | KILL routing tree; keep map for UI only |
| AI MLM v2 (40 Paired Seeds) | Multi-level LINEAGE spawning commissions | Flat TEAM & DIRECT rewards + Script Gate* | No significant LINEAGE gain; risk of spawning collapse; E1u failed gate | Retire SIMPLIFY-TO-DIRECT; enforce Competence Gate |
| Motive Lab MC01 v3 (2,236 states) | Small neural controller (<500MB Gemma/Smol/Julia) | Explicit deterministic rules (M2_rules)* | M2_rules: 6/6; M3_small: 1/6 (positional & early-stop bias) | Fix invariance before any controller training |

## Sources & Methods

- AgentBoard: deterministic offline replay of historical assistant events + synthetic agency-bench comparing organizational topologies A and B′.
- DNS_FOX: Python PoC evaluated across 3 embedding models and 9 tree configurations on a synthetic corpus of 827 memories, 68 resources, and 105 labeled routing queries.
- AI MLM: PyTorch vectorized MARL environments (WORK/WAIT/SPAWN) with REINFORCE + value baseline across 40 pre-registered paired seeds.
