Custom Agent Tactics
Advanced tactics and strategies for maximizing the potential of Claude Code's custom agent system. This guide assumes you understand the fundamentals and focuses on sophisticated usage patterns, community building, and expert-level orchestration.
Agent Chaining & Combinations
Custom agents can be chained together to create sophisticated, reliable workflows:
First use the security-audit agent to identify vulnerabilities,
then use the fix-implementation agent to resolve them,
finally use the test-validation agent to verify the fixes
This approach is particularly powerful for non-coding domains where specialized expertise can be combined:
- UX → SEO → Accessibility analysis pipeline
- Frontend → Accessibility → UX development workflow
- Dumb user → Smart user → UX designer evaluation pipeline
- Research → Analysis → Documentation → Fact checker workflow
- Planning → Implementation → Validation sequence
Effective Agent Combinations
Measure efficacy versus token usage for different combination strategies:
- Minimum viable chain:
25,000tokens for two-agent workflow - Parallel validation: Multiple agents reviewing same artifact
- Sequential refinement: Output from one agent feeds into next
- Multi-part integration: Breaking complex tasks into specialized concerns
Exploring Role Possibilities
The potential for custom agent roles is virtually limitless. Beyond basic specialists, consider experimenting with creative combinations:
Creative and Analytical Roles
- Storytelling specialist, UX empathy mapper, devil's advocate reviewer
- Pattern recognition analyst, edge case identifier, assumption challenger
- Simplicity advocate, complexity reducer, clarity enhancer
Domain-Specific Expertise
- Industry compliance checker, regulatory requirement validator
- Cultural sensitivity reviewer, localization specialist, accessibility advocate
- Performance benchmarker, scalability assessor, maintainability evaluator
Communication and Process Roles
- Technical translator (complex → simple), documentation quality reviewer
- Meeting facilitator, decision synthesizer, conflict resolver
- Project risk assessor, timeline validator, resource optimizer
Experimental Combinations
- Combine multiple perspectives: "Create agents to analyze this from: security + UX + performance perspectives"
- Temporal analysis: "Review this as: past-experience validator + current-state assessor + future-scaling predictor"
- Stakeholder simulation: "Evaluate from: end-user + developer + business-owner viewpoints"
Community Goal: Universal Discipline Coverage
I believe we should be endeavouring as a community to form as many diverse agents as possible to cover ALL disciplines. This includes:
Essential Specialists:
- SEO Agent, Planning Agent, Noob friendly Agent, English Grammar Agent
- Same style Agent, Layout Agent, Design not broken Agent, IP Agent
- MCP tool focused agent, Final check agents
Unexplored Territories:
- Patent search agent, Legal compliance reviewer, Ethics validator
- Accessibility champion, Performance optimizer, Security hardening specialist
- Documentation clarity enforcer, Code simplification advocate
- User journey mapper, Conversion optimizer, A/B testing analyst
The key is continuous experimentation - try unconventional role combinations and measure their effectiveness. Each domain and project may benefit from unique specialist configurations that haven't been explored yet.
Advanced Usage Patterns
Situational Agent Activation
Use --append-system-prompt to provide additional steering for scenarios where specific agents must be invoked:
claude --append-system-prompt "Before any commit, MUST use code-review agent"
This can also be implemented more deterministically via hooks for automated workflow enforcement.
Multi-Chain Workflows
Design workflows where non-destructive analysis happens in parallel, then consolidate findings:
- Research Phase: 3-5 agents investigate different aspects
- Analysis Phase: Specialist agents process research findings
- Implementation Phase: Sequential agents handle implementation steps
- Validation Phase: Quality assurance agents verify results
Benchmarkability Strategy
Custom agents enable granular performance measurement:
- Individual agents can be evaluated independently
- Different inputs can test specific agent capabilities
- Token efficiency can be measured per specialist
- Agents can be tweaked independently instead of modifying master
Claude.md
Advanced Considerations
Information Density vs. Hand-offs
Key Decision: How much information should you pack into a single agent versus creating specialized hand-offs?
Single Agent Approach:
- Lower token cost (one
13kinitialization) - Simpler workflow management
- Risk of reduced specialization
- Harder to benchmark individual capabilities
Specialized Hand-off Approach:
- Higher token cost (multiple
13kinitializations) - Better separation of concerns
- Enhanced benchmarkability
- More granular optimization opportunities
Embracing Weak Models
Explore heavily specialized weak model configurations that have been benchmarked for great performance with lighter models. This could dramatically reduce costs while maintaining quality for specific domains.
Concerns and Limitations
Server Capacity Issues
When servers are busy, multi-chained agent workflows become problematic due to token inefficiency. The 13k initialization cost per agent multiplies quickly, making simple tasks expensive during peak times.
Activation Reliability
Monitor how reliably Claude routes tasks to appropriate agents, especially when you have many custom agent options. Too many choices can reduce routing accuracy.
Token Efficiency Trade-offs
The minimum cost for agent chaining (25,000+ tokens) merits caution. Balance the sophistication of multi-agent workflows against token costs, especially for routine tasks.
Community Vision
The true potential of custom agents lies in collaborative development:
Agent Sharing Economy
- Refine agents across multiple projects
- Share battle-tested configurations
- Build upon community expertise
- Create domain-specific agent libraries
Quality Standards
- Benchmark agent performance metrics
- Document effective usage patterns
- Establish best practices for different domains
- Measure token efficiency vs. value delivered
Ecosystem Growth
Think of custom agents like Pokémon - collect diverse specialists, refine their capabilities, and combine them strategically. The goal is building a comprehensive ecosystem where any task can be handled by the appropriate specialist combination.
Community Collaboration Strategy
Sharing Mechanisms:
- Standardized agent configuration templates
- Performance benchmarking frameworks
- Community repositories of battle-tested agents
- Cross-project compatibility standards
Measurement and Iteration:
- Token cost vs. value analysis
- Reliability metrics for different domains
- Effectiveness comparison across similar agents
- Community feedback loops for continuous improvement
Collective Intelligence:
- Domain expert networks contributing specialized agents
- Peer review processes for agent quality
- Collaborative debugging of agent behavior
- Knowledge transfer through agent documentation
Expert-Level Tactics
Advanced Orchestration
Context Management:
- Strategic agent selection to preserve main thread context
- Intelligent consolidation of multi-agent findings
- Context window optimization through agent delegation
Workflow Design:
- Dependency mapping for sequential agent chains
- Parallel processing optimization for independent tasks
- Error handling and fallback strategies in agent chains
Performance Optimization:
- Agent warm-up strategies for frequently used specialists
- Batch processing techniques for similar tasks
- Load balancing across different agent types
Master the art of agent orchestration by treating custom agents as specialized team members. Design workflows that leverage their strengths while minimizing token costs through strategic sequencing and parallel processing.
See Also: Agent Engineering | Custom Agents | Split Role Sub-Agents | Hooks | Sanity Check
InventorBlack