Multi-Agent Parallel Scaling
Relevant source files
The following files were used as context for generating this wiki page:
Multi-Agent Parallel Scaling in the FusionOps environment refers to the architectural capability of the Hermes agent system to execute multiple specialized autonomous agents simultaneously. This system is designed to handle high-volume intelligence tasks—such as brand protection, risk analysis, and account monitoring—by distributing workloads across parallel processing phases before converging for final decision-making.
The system utilizes a "Phase-based" execution model where a team of primary agents (Phase 1) gathers data in parallel, followed by serial execution (Phase 2) where an aggregator agent synthesizes the results into actionable verdicts. This approach reduces total execution time from hours to minutes, typically completing nightly analysis pipelines for numerous accounts within 2 to 5 minutes.
Pipeline Architecture
The nightly analysis pipeline is the primary implementation of parallel scaling within FusionOps. It is triggered by the /fbis-run skill and utilizes four distinct agents running concurrently to perform multi-dimensional risk assessment.
Parallel Execution Phase (Phase 1)
In this phase, the system scales horizontally by initiating four specialized "Ban Intelligence" agents at the same time. These agents operate independently to query different data points from the FBIS MCP (Model Context Protocol) Server and D1 databases.
- ARGUS: Analyzes proxy and IP risk.
- NEXUS: Evaluates account isolation scoring.
- IRIS: Conducts traffic quality analysis.
- CHRONO: Predicts ban timelines based on historical patterns.
Sources: HERMES_FBIS_AUDIT.txt:62-73, HERMES_FBIS_AUDIT.txt:82-87
Synthesis Phase (Phase 2)
Scaling transitions from parallel to serial execution once all Phase 1 agents complete their tasks. The VERDICT agent acts as a risk aggregator, pulling the weighted scores from the four parallel agents to generate a final verdict_score (0-100).
Sources: HERMES_FBIS_AUDIT.txt:89-93, hermes/skills/verdict.skill.md
Pipeline Flow Diagram
The following diagram illustrates the transition from parallel data gathering to serialized aggregation and alerting.
The diagram shows how the system scales out during the data gathering phase and scales in for the final verdict.
Sources: HERMES_FBIS_AUDIT.txt:82-100
Scaling via Cron Orchestration
Parallel scaling is also achieved through scheduled orchestration within the config.yaml file. The Hermes Cron Scheduler allows for different agents to be assigned specific schedules, ensuring that specialized monitoring tasks (like brand-jack detection) happen frequently without blocking longer-term discovery tasks.
Scheduled Agent Team
The Hermes team for trademark monitoring consists of four agents with varying frequencies, managed by the core runtime scheduler.
| Agent | Purpose | Schedule | Scaling Frequency |
|---|---|---|---|
| AEGIS | Brand-Jack Detection | 0 */2 * * * | Every 2 hours |
| SCOUT | Competitor Ad Copy | 0 6 * * * | Daily |
| HERALD | Trend & Reputation | 0 8 * * * | Daily |
| ORACLE | Keyword Discovery | 0 9 * * 1 | Weekly |
Sources: hermes-agents/INDEX.md:5-30, infra/hermes/config.yaml
Implementation in Configuration
The config.yaml file defines the parallel capabilities by enabling multiple concurrent skill entries under the crons key. Each entry can point to a different .env file or shared environment variables to maintain isolation during execution.
crons:
aegis-brand-jack:
schedule: "0 */2 * * *"
skill: "trademark-monitoring/aegis"
enabled: true
env_file: /opt/fusionops/agents/.env
scout-ad-copy:
schedule: "0 6 * * *"
skill: "trademark-monitoring/scout"
enabled: true
Sources: hermes-agents/DEPLOYMENT_CHECKLIST.md:65-80, infra/hermes/config.yaml
Risk Scoring and Automated Responses
Parallel scaling necessitates a unified scoring system to process the concurrent streams of data. Each agent provides a component score which is then processed through weighted averaging.
Risk Component Mapping
The risk system scales its response based on the aggregated results of the parallel agents.
- Healthy (0-30): Safe; no action required.
- Watch (31-60): Monitor closely.
- Risk (61-80): Action required (e.g., 50% budget reduction).
- Critical (81-100): Pause account immediately.
Automated Pause Rules Sequence
When a critical threshold is met, the system executes automated pause rules across the integrated Cloudflare and API environment.
The sequence shows the transition from multi-agent data input to a singular automated decision.
Sources: HERMES_FBIS_AUDIT.txt:111-131
Scaling Efficiency and Optimization
To support a large number of agents running in parallel, the system implements optimization plugins to reduce token consumption and latency.
Ecosystem Plugins
- Disk-cleanup: Prevents state files from consuming excessive VPS storage.
- Caveman Compression: Reduces token usage by up to 80% per call.
- Graphify: Visualizes agent dependencies and MCP connections on port 8767.
- Mission Control: A dedicated orchestration dashboard (port 3001) for monitoring parallel agent health.
Sources: HERMES_FBIS_AUDIT.txt:234-245, infra/hermes/config.yaml
Resource Usage Estimates
A scaled 4-agent team consumes approximately 2,100 API calls per month, representing only 21% of the standard free tier for external intelligence APIs (like DataForSEO). This efficiency allows the system to scale to even larger agent teams without significant cost increases.
Sources: hermes-agents/BUILD_SUMMARY.md:155-165
Summary
The Multi-Agent Parallel Scaling feature in FusionOps provides a robust framework for autonomous operations. By utilizing a phase-based pipeline, the system allows ARGUS, NEXUS, IRIS, and CHRONO to gather intelligence concurrently, while the VERDICT agent ensures a centralized point of governance. This architecture, supported by cron orchestration and token compression, allows the platform to maintain high-frequency monitoring and rapid automated response times across hundreds of advertising accounts.