Specialist Agents

DojOps includes 32 built-in specialist agents for keyword-based prompt routing, plus support for user-defined custom agents. Each agent has a domain-specific system prompt, keyword set, and optional tool dependencies.

How Routing Works

When you send a prompt to DojOps, the AgentRouter scores it against each agent’s keyword list:

1. Keyword matching, Each agent has a set of domain-specific keywords. The router counts how many keywords appear in the prompt.
2. Primary keyword boosting, Each agent also has a primaryKeywords subset (typically 3-5 keywords) that receive higher weight during scoring, ensuring strong domain signals are prioritized.
3. Confidence scoring, The score is normalized based on keyword match density. Higher scores indicate stronger domain relevance.
4. Threshold check, If the highest-scoring agent exceeds the confidence threshold, the prompt is routed to that specialist.
5. Fallback, If no agent exceeds the threshold, the prompt goes to the general-purpose DevOpsAgent.

The selected agent’s system prompt is prepended to the LLM request, providing domain-specific instructions and constraints.

Agent List

Keywords Reference

Each agent is matched by the following keyword sets:

Tool Dependencies

Some agents declare external tool dependencies that enhance their capabilities:

Using Agents

CLI

# List all agents
dojops agents list
 
# Show agent details (partial names supported)
dojops agents info terraform              # matches terraform-specialist
dojops agents info security               # matches security-auditor
dojops agents info cloud                  # matches cloud-architect
dojops agents info terraform-specialist   # exact name also works
 
# Pin chat to an agent
dojops chat --agent=terraform

API

# List all agents
curl http://localhost:3000/api/agents

Automatic Routing

Agents are selected automatically based on prompt content. No manual routing is needed:

# Routes to terraform-specialist (matches: terraform, s3, iac)
dojops "Create a Terraform config for S3"
 
# Routes to kubernetes-specialist (matches: kubernetes, deployment, nginx)
dojops "Write a Kubernetes deployment for nginx"
 
# Routes to cicd-specialist (matches: github actions, pipeline, ci)
dojops "Set up GitHub Actions CI pipeline"
 
# Routes to ops-cortex (matches: plan, multi-step, end-to-end)
dojops plan "Set up end-to-end CI/CD with Docker and Kubernetes"

Custom Agents

In addition to the 32 built-in agents, you can create your own custom agents. Custom agents participate in the same keyword-based routing as built-in agents and can even override built-in agents by name.

Agent Definition Format

Each custom agent is a directory with a structured README.md:

.dojops/agents/sre-specialist/README.md

# SRE Specialist
 
## Domain
 
site-reliability
 
## Description
 
SRE specialist for incident response, reliability engineering, and observability.
 
## System Prompt
 
You are an SRE specialist. You specialize in:
 
- Incident response and post-mortems
- SLO/SLI design and error budgets
- Chaos engineering and resilience testing
- On-call runbooks and escalation procedures
- Capacity planning and performance optimization
 
When asked about infrastructure, focus on reliability patterns...
 
## Keywords
 
sre, incident, reliability, error budget, slo, chaos, postmortem, runbook, on-call, resilience

Required sections: ## Domain, ## Description, ## System Prompt, ## Keywords (comma-separated).

Discovery Paths

Custom agents are discovered from two locations:

Project agents override global agents with the same name.

Creating Custom Agents

LLM-generated (recommended):

dojops agents create "an SRE specialist for incident response and reliability"

The LLM generates a complete agent definition (name, domain, description, system prompt, keywords) and writes the README.md to .dojops/agents/<name>/.

Manual creation:

dojops agents create --manual

Interactive prompts guide you through defining name, domain, description, system prompt, and keywords.

Global agents (shared across projects):

dojops agents create --global "a cost optimization specialist"

Managing Custom Agents

# List all agents (built-in + custom)
dojops agents list
 
# Show agent details (partial names supported, includes source path for custom agents)
dojops agents info sre                    # matches sre-specialist
 
# Remove a custom agent
dojops agents remove sre-specialist

Routing with Custom Agents

Custom agents are routed exactly like built-in agents, by keyword matching. If a custom agent’s keywords match the prompt with higher confidence than any built-in agent, the custom agent handles the request:

# Routes to custom sre-specialist (matches: sre, error budget, slo)
dojops "Design SLOs and error budgets for our payment service"

Agents in Plan Execution

Beyond prompt routing, agents also participate in plan execution. When dojops plan decomposes a goal into a task graph, the LLM assigns a specialist agent to each task based on domain relevance.

During execution, the assigned agent’s system prompt is injected as domain context into the skill’s LLM call. This means the LLM sees both:

• Agent expertise, Domain-specific knowledge (e.g., Terraform state management, K8s security best practices)
• Skill instructions, The skill’s generation format and rules (e.g., “output valid HCL”)

Goal: "Set up CI/CD with Docker and Kubernetes"

Task 1: github-actions  [cicd-specialist]       -> CI/CD pipeline expertise
Task 2: dockerfile      [docker-specialist]     -> Multi-stage build patterns
Task 3: kubernetes      [kubernetes-specialist]  -> Deployment best practices

Agent assignment is optional, tasks where no specialist clearly matches proceed without additional domain context. Both built-in and custom agents participate in delegation.

See Task Planner for details on the decomposition and execution flow.

Built-in Agent Configuration

Built-in agents are defined in packages/core/src/agents/specialists.ts. Each agent specifies:

• name, Unique identifier
• domain, Category label
• description, System prompt context
• keywords, Array of matching keywords for routing
• primaryKeywords, Subset of keywords with boosted weight for stronger routing signals
• toolDependencies, Optional external tools

All agents include a NO_FOLLOWUP_INSTRUCTION suffix ensuring single-shot responses without conversational follow-ups.

Specialized Analyzers

In addition to the 32 routed agents, DojOps provides three specialized analyzers that are invoked directly (not via AgentRouter):

DevOps Checker

The DevOpsChecker (packages/core/src/agents/devops-checker.ts) analyzes DevOps files detected during dojops init for quality, security, and best practices. It produces:

• Maturity score (0-100), Minimal (0-25), Basic (26-50), Good (51-75), Excellent (76-100)
• Findings, Severity-ranked issues (critical, error, warning, info) categorized as security, quality, best-practice, performance, or reliability
• Missing files, Important DevOps files the project should have but doesn’t

dojops check                  # Display formatted report
dojops check --output json    # Machine-readable JSON output