Agent Types Reference

February 21, 2026 ยท View on GitHub

Complete definitions and capabilities for all 41 specialized agent types.


Overview

Loki Mode has 41 predefined agent types organized into 8 specialized swarms (37 domain agents + 4 orchestration agents). The orchestrator spawns only the agents needed for your project -- typically 5-10 for simple projects, more for complex ones.


Engineering Swarm (8 types)

AgentCapabilities
eng-frontendReact/Vue/Svelte, TypeScript, Tailwind, accessibility, responsive design, state management
eng-backendNode/Python/Go, REST/GraphQL, auth, business logic, middleware, validation
eng-databasePostgreSQL/MySQL/MongoDB, migrations, query optimization, indexing, backups
eng-mobileReact Native/Flutter/Swift/Kotlin, offline-first, push notifications, app store prep
eng-apiOpenAPI specs, SDK generation, versioning, webhooks, rate limiting, documentation
eng-qaUnit/integration/E2E tests, coverage, automation, test data management
eng-perfProfiling, benchmarking, optimization, caching, load testing, memory analysis
eng-infraDocker, K8s manifests, IaC review, networking, security hardening

Operations Swarm (8 types)

AgentCapabilities
ops-devopsCI/CD pipelines, GitHub Actions, GitLab CI, Jenkins, build optimization
ops-sreReliability, SLOs/SLIs, capacity planning, on-call, runbooks
ops-securitySAST/DAST, pen testing, vulnerability management, security reviews
ops-monitorObservability, Datadog/Grafana, alerting, dashboards, log aggregation
ops-incidentIncident response, runbooks, RCA, post-mortems, communication
ops-releaseVersioning, changelogs, blue-green, canary, rollbacks, feature flags
ops-costCloud cost optimization, right-sizing, FinOps, reserved instances
ops-complianceSOC2, GDPR, HIPAA, PCI-DSS, audit preparation, policy enforcement

Business Swarm (8 types)

AgentCapabilities
biz-marketingLanding pages, SEO, content, email campaigns, social media
biz-salesCRM setup, outreach, demos, proposals, pipeline management
biz-financeBilling (Stripe), invoicing, metrics, runway, pricing strategy
biz-legalToS, privacy policy, contracts, IP protection, compliance docs
biz-supportHelp docs, FAQs, ticket system, chatbot, knowledge base
biz-hrJob posts, recruiting, onboarding, culture docs, team structure
biz-investorPitch decks, investor updates, data room, cap table management
biz-partnershipsBD outreach, integration partnerships, co-marketing, API partnerships

Data Swarm (3 types)

AgentCapabilities
data-mlModel training, MLOps, feature engineering, inference, model monitoring
data-engETL pipelines, data warehousing, dbt, Airflow, data quality
data-analyticsProduct analytics, A/B tests, dashboards, insights, reporting

Product Swarm (3 types)

AgentCapabilities
prod-pmBacklog grooming, prioritization, roadmap, specs, stakeholder management
prod-designDesign system, Figma, UX patterns, prototypes, user research
prod-techwriterAPI docs, guides, tutorials, release notes, developer experience

Growth Swarm (4 types)

AgentCapabilities
growth-hackerGrowth experiments, viral loops, referral programs, acquisition
growth-communityCommunity building, Discord/Slack, ambassador programs, events
growth-successCustomer success, health scoring, churn prevention, expansion
growth-lifecycleEmail lifecycle, in-app messaging, re-engagement, onboarding

Review Swarm (3 types)

AgentCapabilities
review-codeCode quality, design patterns, SOLID, maintainability, best practices
review-businessRequirements alignment, business logic, edge cases, UX flows
review-securityVulnerabilities, auth/authz, OWASP Top 10, data protection

Orchestration Swarm (4 types)

Source: Cursor Scaling Learnings - patterns proven at large agent scale

AgentCapabilities
orch-plannerTask decomposition, dependency analysis, work distribution, sub-planner spawning
orch-sub-plannerDomain-specific planning (frontend, backend, testing), recursive task breakdown
orch-judgeCycle continuation decisions, goal completion assessment, escalation triggers
orch-coordinatorCross-stream coordination, merge decisions, conflict resolution

Recursive Sub-Planner Pattern

Main Planner (orch-planner)
    |
    +-- Sub-Planner Frontend (orch-sub-planner)
    |       +-- eng-frontend (Component A)
    |       +-- eng-frontend (Component B)
    |
    +-- Sub-Planner Backend (orch-sub-planner)
    |       +-- eng-backend (API)
    |       +-- eng-database (Schema)
    |
    +-- Sub-Planner Testing (orch-sub-planner)
            +-- eng-qa (Unit tests)
            +-- eng-qa (E2E tests)

Benefits:

  • Planning scales horizontally (not bottlenecked on single planner)
  • Each sub-planner has focused domain context
  • Enables true parallel planning

Judge Agent Protocol

judge_agent:
  trigger: After milestone OR every N iterations
  inputs:
    - current_state: ".loki/state/orchestrator.json"
    - original_goal: "PRD requirements"
    - recent_progress: "Last 10 completed tasks"
    - resource_consumption: "Tokens, time, iterations"
  outputs:
    - CONTINUE: "More work needed toward goal"
    - COMPLETE: "Goal achieved, finalize"
    - ESCALATE: "Human intervention required"
    - PIVOT: "Change approach, current path blocked"
  model: haiku  # Fast decision, low cost

Agent Execution Model

Claude Code does NOT support background processes. Agents execute via:

  1. Role Switching (Recommended): Orchestrator maintains agent queue, switches roles per task
  2. Sequential: Execute agents one at a time (simple, reliable)
  3. Parallel via tmux: Multiple Claude Code sessions (complex, faster)
# Option 1: Sequential (simple, reliable)
for agent in frontend backend database; do
  claude -p "Act as $agent agent..." --dangerously-skip-permissions
done

# Option 2: Parallel via tmux (complex, faster)
tmux new-session -d -s loki-pool
for i in {1..5}; do
  tmux new-window -t loki-pool -n "agent-$i" \
    "claude --dangerously-skip-permissions -p '$(cat .loki/prompts/agent-$i.md)'"
done

# Option 3: Role switching (recommended)
# Orchestrator maintains agent queue, switches roles per task

Model Selection by Agent Type

Task TypeModelReason
Planning/ArchitectureOpusDeep reasoning for design decisions
ImplementationSonnetDevelopment workload
Code ReviewSonnetBalanced quality/cost
Security ReviewSonnetBalanced quality/cost
Business Logic ReviewSonnetBalanced quality/cost
DocumentationSonnetStraightforward writing
Unit tests/Quick fixesHaikuFast iteration

Agent Lifecycle

SPAWN -> INITIALIZE -> POLL_QUEUE -> CLAIM_TASK -> EXECUTE -> REPORT -> POLL_QUEUE
           |              |                        |          |
           |         circuit open?             timeout?    success?
           |              |                        |          |
           v              v                        v          v
     Create state    WAIT_BACKOFF              RELEASE    UPDATE_STATE
                          |                    + RETRY         |
                     exponential                              |
                       backoff                                v
                                                    NO_TASKS --> IDLE (5min)
                                                                    |
                                                             idle > 30min?
                                                                    |
                                                                    v
                                                               TERMINATE

Dynamic Scaling Rules

ConditionActionCooldown
Queue depth > 20Spawn 2 agents of bottleneck type5min
Queue depth > 50Spawn 5 agents, alert orchestrator2min
Agent idle > 30minTerminate agent-
Agent failed 3x consecutiveTerminate, open circuit breaker5min
Critical task waiting > 10minSpawn priority agent1min
Circuit breaker half-openSpawn 1 test agent-
All agents of type failedHALT, request human intervention-

Agent Context Preservation

Lineage Rules

  1. Immutable Inheritance: Agents CANNOT modify inherited context
  2. Decision Logging: All decisions MUST be logged to agent context file
  3. Lineage Reference: All commits MUST reference parent agent ID
  4. Context Handoff: When agent completes, context is archived but lineage preserved

Preventing Context Drift

  1. Read .agent/sub-agents/${parent_id}.json before spawning
  2. Inherit immutable context (tech stack, constraints, decisions)
  3. Log all new decisions to own context file
  4. Reference lineage in all commits
  5. Periodic context sync: check if inherited context has been updated upstream