1,110 Agent Definitions (~70 runtime): Building an AI Workforce

Prismatic Platform operates 552 autonomous agents. Not 552 copies of the same agent – 552 uniquely specialized agents, each with a defined purpose, toolset, and behavioral constraints. This post explains how we built and manage this agent ecosystem.

#What Is an Agent?

In Prismatic, an agent is a markdown-defined specification that describes:

# Agent: czech-ares-analyst

## Purpose
Analyze Czech business registry (ARES) data for due diligence investigations.

## Capabilities
- Query ARES API for company details
- Cross-reference with Justice.cz records
- Verify ICO, address, and director information
- Generate structured company profiles

## Tools
- ARES API adapter
- Justice.cz adapter
- Entity resolution module
- Report generator

## Constraints
- Maximum 30 API calls per minute
- Must verify data against at least 2 sources
- All findings must include confidence scores

Agents are specifications, not implementations. The AIAD (AI-Assisted Development) runtime interprets these specifications and coordinates execution using available tools and modules.

#Agent Categories

The 1,110 agent definitions (~70 runtime) span 12 categories:

#Category distribution (interactive)

<div class=”h-80 w-full rounded-xl border border-gray-800 bg-gray-950”

 data-chart-type="doughnut"
 data-chart-data="agents.categories"
 data-chart-options='{"responsive":true,"maintainAspectRatio":false,"plugins":{"legend":{"position":"right","labels":{"color":"#cbd5e1","font":{"size":11}}}}}'></div>

#The 552-agent mesh, rendered in 3D

<div class=”h-96 w-full rounded-xl border border-gray-800 bg-gray-950 three-container”

 data-three-scene="network-graph"
 data-three-nodes="60"
 data-three-edges="120"></div>

Each node represents a specialized agent; edges show coordination channels through the AIAD orchestrator.

#Self-Registration

Agents register themselves at startup. Each agent definition in .aiad/agents/ is discovered and indexed:

defmodule Prismatic.Agents.Registry do
  @table :agent_registry

  def discover_and_register do
    agent_dir = Path.join([:code.priv_dir(:prismatic), "..", "..", ".aiad", "agents"])

    agent_dir
    |> Path.join("*.agent.md")
    |> Path.wildcard()
    |> Enum.each(fn path ->
      agent = parse_agent_definition(path)
      :ets.insert(@table, {agent.slug, agent})
    end)
  end
end

The ETS-backed registry provides sub-microsecond lookups. Adding a new agent requires only creating a markdown file – no code changes, no configuration updates.

#Agent Coordination

Complex tasks require multiple agents working together. The orchestration system manages this:

# Multi-agent DD investigation
PrismaticAgents.orchestrate(%{
  mission: "Comprehensive DD on Target s.r.o.",
  agents: [
    {:czech_ares_analyst, %{entity: "Target s.r.o.", ico: "12345678"}},
    {:sanctions_screener, %{entity: "Target s.r.o."}},
    {:financial_analyst, %{entity: "Target s.r.o."}},
    {:director_background, %{names: ["Jan Novak", "Eva Svobodova"]}}
  ],
  strategy: :parallel,
  timeout: 300_000
})

The orchestrator:

1. Validates that all requested agents exist and are available
2. Launches agents in parallel (or sequentially if dependencies exist)
3. Collects results from all agents
4. Merges findings using entity resolution
5. Produces a unified report

#Agent Evolution

Agents evolve alongside the platform. The evolution system:

• Tracks agent performance – success rates, execution times, user feedback
• Identifies underperforming agents – agents with low success rates or redundant capabilities
• Proposes improvements – updated specifications based on performance data
• Removes duplicates – agents with identical capabilities are consolidated

Gen 19 reduced the agent population from 532 to 530 through targeted deduplication. The count has since grown to 552 as new capabilities were added.

#Agent Quality Standards

Every agent must meet quality criteria:

1. Unique purpose – no two agents should have identical capabilities
2. Defined constraints – rate limits, data access boundaries, confidence thresholds
3. Tool specification – explicit list of tools the agent can use
4. Error handling – defined behavior for failures and edge cases
5. Documentation – clear description of capabilities and limitations

The agent registry validates these criteria at registration time. Agents that fail validation are flagged for review.

#Special Operations Agents

The platform includes 85 special operations agents inspired by military intelligence units:

• Navy SEAL – deep-water OSINT operations in hostile digital environments
• Delta Force – precision OSINT operations with surgical accuracy
• Ghost Recon – stealth reconnaissance with maximum operational security
• Falcon Strike – rapid multi-vector intelligence assault
• Siege Master – long-term intelligence siege with systematic analysis

These agents are tuned for specific OSINT scenarios: comprehensive target profiling, time-sensitive intelligence gathering, or sustained monitoring campaigns.

#The AIAD Standard

AIAD (AI-Assisted Development) is the standard that governs agent behavior:

.aiad/
├── agents/          # 552 agent definitions
├── commands/        # 228 command specifications
├── doctrine/        # Behavioral constraints
├── policies/        # Quality and security policies
├── prompts/         # Reusable prompt templates
└── workflows/       # Multi-step workflow definitions

The standard ensures consistency across all agents while allowing domain-specific specialization.

#Scaling Considerations

Managing 1,110 agent definitions (~70 runtime) requires infrastructure:

• Registry performance – ETS-backed, sub-microsecond lookups
• Agent discovery – file-system scanning at startup, cached in ETS
• Coordination overhead – orchestrator manages concurrent agent execution
• Memory footprint – agent definitions are lightweight (markdown + metadata)
• Monitoring – telemetry tracks agent invocations, durations, and outcomes

The agent count is tracked dynamically by mix prismatic.stats.validate – never hardcoded.

#Conclusion

1,110 agent definitions (~70 runtime) is not a vanity metric. It represents the platform’s accumulated intelligence about how to perform specific tasks in specific domains. Each agent encodes expertise that would otherwise live in documentation, tribal knowledge, or individual memory. By formalizing this expertise as agent specifications, we make it discoverable, composable, and evolvable.

Browse all 1,110 agent definitions (~70 runtime) at Agent Hub or learn about Agent Coordination for orchestration patterns.