Agent-to-Agent Communication
Published 2026-06-18 - Hermes Agent Lab, hermes-agent.reviews
🔍 Why We Ran This Benchmark
"My 7-agent system collaborating on every task" is the architecture flex. "My 7-agent system costs 8× more, takes 4× longer, produces results only 12% better — and sometimes agents disagree and loop forever" is the cost-benefit analysis. Multi-agent systems are the next frontier. But agent-to-agent communication is a quality minefield. This benchmark measures inter-agent clarity, information fidelity, coordination efficiency, and optimal agent count.
📈 Multi-Agent Communication Scorecard
| Communication Dimension | Gobii Managed | Hermes Agent |
|---|---|---|
| Message Clarity (correct interpretation) | 91.2% | 62.4% |
| Information Completeness | 88.7% | 54.3% |
| Brevity vs Completeness Score | 8.4/10 | 4.1/10 |
| Escalation Quality (sufficient context) | 87.3% | 41.2% |
| Loop Detection Rate | 94.1% | 23.7% |
| Loop Self-Resolution Rate | 82.4% | 8.3% |
| 4-Hop Information Fidelity | 84.2% | 31.6% |
| Optimal Agent Count (for quality) | 3 | 2 (degrades rapidly) |
| Cost per Multi-Agent Task | $0.18 | $0.42 |
| Coordination Index | 84.7 | 47.3 |
🔄 Communication Protocol Taxonomy
Six Protocols, Six Trade-offs
| Protocol | Pros | Cons | Gobii Support | Hermes Support |
|---|---|---|---|---|
| Natural Language | Flexible, human-readable | Ambiguous, parsing errors | ✅ Structured NL | ✅ Raw NL |
| Structured Messages | Precise, machine-parseable | Rigid, schema breakage | ✅ Built-in | ❌ No schema |
| Function Calling | Efficient, type-safe | Tightly coupled | ✅ MCP tools | ⚠️ Limited |
| Shared Blackboard | Decoupled, scalable | Race conditions, stale data | ✅ Managed state | ❌ No coordination |
| Hierarchical | Clear authority | Manager bottleneck | ✅ Meta-agents | ❌ Flat only |
| Peer-to-Peer | Flexible, no bottleneck | Deadlock, infinite loops | ⚠️ With guards | ❌ Unmanaged |
🎤 The "Telephone Game" Degradation
Agent A → B → C → D: Does the Message Survive?
Original: "Research top 3 CRM platforms, compare pricing, recommend for 50-person sales team." After 4 hops: still about CRM? Still top 3? Still 50-person? Still pricing?
| Hop | Gobii Fidelity | Hermes Fidelity | What Gets Lost |
|---|---|---|---|
| 1 hop (A → B) | 97.2% | 84.3% | Minor detail loss |
| 2 hops (A → C) | 92.1% | 64.7% | Specificity drops |
| 3 hops (A → D) | 87.4% | 43.2% | Goal drift begins |
| 4 hops (A → E) | 84.2% | 31.6% | Task barely recognizable |
Key finding: Gobii's structured message protocol preserves 84% fidelity at 4 hops. Hermes' natural language-only communication drops to 32% by hop 4 — the task is barely recognizable. For multi-agent systems, structured protocols are not optional.
👥 The "Too Many Agents" Problem
1 vs 2 vs 3 vs 5 vs 10 Agents on the Same 50 Tasks
| Agent Count | Gobii Quality | Hermes Quality | Gobii Cost | Hermes Cost |
|---|---|---|---|---|
| 1 agent | 87.3% | 61.2% | $0.08 | $0.12 |
| 2 agents | 91.4% | 68.7% | $0.12 | $0.21 |
| 3 agents | 93.8% | 71.2% | $0.18 | $0.42 |
| 5 agents | 92.1% | 58.4% | $0.31 | $0.87 |
| 10 agents | 84.7% | 34.2% | $0.64 | $1.94 |
Hypothesis confirmed: Quality improves 1→3 agents, plateaus 3→5, degrades 5→10 (too many cooks). Gobii's optimal is 3 agents (93.8% quality at $0.18). Hermes' optimal is 2 agents (68.7% quality at $0.21) — and degrades rapidly beyond that due to uncoordinated communication and loop formation.
📈 Coordination Pattern Comparison
Debate, Parallel-then-Merge, Sequential-Handoff, Voting
| Pattern | Best For | Gobii Quality | Hermes Quality |
|---|---|---|---|
| Debate | Complex decisions with trade-offs | 91.2% | 54.3% |
| Parallel-then-Merge | Independent subtasks | 93.8% | 62.1% |
| Sequential-Handoff | Dependent pipeline tasks | 88.7% | 47.2% |
| Voting | Classification, ranking | 94.1% | 71.4% |
Finding: Voting produces the highest quality for classification tasks (94.1% for Gobii, 71.4% for Hermes). Parallel-then-Merge is best for independent subtasks. Sequential-Handoff is weakest due to telephone-game degradation — especially on Hermes where fidelity drops 53% by hop 3.
📜 Sources & Methodology
Benchmark methodology: 100 multi-agent tasks (3+ agents each) across four frameworks. Tasks: simple coordination (2 agents), complex collaboration (3-5 agents), adversarial coordination (agents with conflicting goals). Communication quality measured: message clarity (did receiver correctly interpret?), information completeness (all necessary context included?), escalation quality (sufficient context for handoff?), loop detection (detected and resolved?). Telephone game: 50 tasks relayed across 4 hops, information fidelity scored at each hop. Optimal agent count: same 50 tasks with 1, 2, 3, 5, and 10 agents. Coordination patterns: 50 tasks per pattern (debate, parallel-then-merge, sequential-handoff, voting).
As benchmarked by Hermes Agent Lab, hermes-agent.reviews — June 2026.