Workflow Customization

Weave's agent system is fully configurable. Loom remains the main user-facing orchestrator, but you can disable agents you don't need, add custom specialists, change which models back each agent, and reshape the delegation workflow. All of this is done through your .weave config.

This page shows practical recipes for common workflow customizations.

Understanding the Default Workflow

Out of the box, Weave uses this delegation flow:

User → Loom (orchestrator)
         ├── Pattern (planning)
         ├── Thread (exploration)
         ├── Spindle (research)
         ├── Shuttle (coding)
         ├── Weft (review)
         └── Warp (security)
         
Plan execution via `/start-work`: Loom → Tapestry (coordinates step-by-step delegation to Shuttle)

Every part of this is customizable:

• Disable any agent you don't use
• Reconfigure any agent's model, temperature, or tool policy
• Add new agents that participate in the workflow
• Reshape delegation by adding triggers that tell Loom when to use each agent

Recipe: Dual Security Review

Get two independent security perspectives by adding a second reviewer backed by a different model:

# Keep the built-in Warp (Anthropic-backed) and add a GPT-backed reviewer
agent security-gpt {
  description "Independent security review using OpenAI for diverse coverage"
  prompt "You are a security auditor. Review code for OWASP Top 10 vulnerabilities, authentication flaws, injection attacks, data exposure, and insecure dependencies. Be thorough and skeptical. Produce a clear APPROVE or REJECT verdict."
  models ["openai/gpt-5"]
  mode subagent
  temperature 0.2

  tool_policy {
    read allow
    write deny
    execute deny
    delegate deny
  }

  triggers [
    { domain "Security" trigger "Independent security review using a different model for diverse analysis" }
  ]
}

You can instruct Loom to use both reviewers via prompt_append:

agent loom {
  prompt_append "When reviewing security-sensitive changes, delegate to BOTH Warp AND Security-GPT for independent reviews."
}

Recipe: Minimal Agent Setup

Strip Weave down to just the agents you need. For example, a solo developer who doesn't need planning or research:

disable agents ["pattern", "spindle", "weft", "warp"]

agent loom {
  prompt_append "Skip planning phases. Work directly with Thread for exploration and Shuttle for coding."
}

This leaves only loom, tapestry, shuttle, and thread, a lean setup for rapid prototyping.

Recipe: Replace the Review Pipeline

Swap out the built-in reviewers for custom ones tailored to your team's standards:

# Disable built-in reviewers
disable agents ["weft", "warp"]

# Add team-specific reviewers
agent code-reviewer {
  description "Reviews code against team standards and patterns"
  prompt "Review code changes for correctness, maintainability, and adherence to team conventions. Check test coverage. Approve or request changes with specific, actionable feedback."
  models ["anthropic/claude-sonnet-4-5"]
  mode subagent
  skills ["company-standards", "testing"]

  tool_policy {
    read allow
    write deny
    execute deny
    delegate deny
  }

  triggers [
    { domain "Code Review" trigger "Code quality review, standards compliance, test coverage verification" }
  ]
}

agent compliance-checker {
  description "Checks for regulatory and licensing compliance"
  prompt "Check code for licensing compliance, PII handling, GDPR requirements, and regulatory concerns. Flag any issues with third-party dependencies."
  models ["anthropic/claude-sonnet-4-5"]
  mode subagent

  tool_policy {
    read allow
    write deny
    execute deny
    delegate deny
  }

  triggers [
    { domain "Compliance" trigger "License checks, PII handling, regulatory review" }
  ]
}

Recipe: Domain-Specialized Specialist Agents

Instead of one general-purpose Shuttle, add domain-specific specialist agents:

agent frontend-dev {
  description "React/TypeScript frontend specialist"
  prompt_file "frontend.md"
  models ["anthropic/claude-sonnet-4-5"]
  mode subagent
  skills ["react-best-practices", "typescript-strict", "accessibility"]

  triggers [
    { domain "Frontend" trigger "React components, CSS, accessibility, UI/UX implementation" }
  ]
}

agent backend-dev {
  description "Python FastAPI backend specialist"
  prompt_file "backend.md"
  models ["anthropic/claude-sonnet-4-5"]
  mode subagent
  skills ["python-patterns", "api-design"]

  triggers [
    { domain "Backend" trigger "API endpoints, database queries, business logic, background jobs" }
  ]
}

agent infra-dev {
  description "DevOps and infrastructure specialist"
  models ["anthropic/claude-sonnet-4-5"]
  mode subagent
  skills ["docker", "ci-cd"]

  triggers [
    { domain "Infrastructure" trigger "Docker, CI/CD pipelines, deployment, monitoring, IaC" }
  ]
}

Loom's delegation table will now include all three specialists, and it will route tasks based on the domain triggers.

Recipe: Research-Heavy Workflow

For teams that need extensive research before coding, add a specialized research agent:

agent api-researcher {
  description "Researches third-party APIs and libraries"
  prompt "Research third-party APIs, SDKs, and libraries. Read documentation, find examples, and summarize integration approaches. Never write production code, only research and recommend."
  models ["anthropic/claude-sonnet-4-5"]
  mode subagent

  tool_policy {
    read allow
    write deny
    execute deny
    delegate deny
    network allow
  }

  triggers [
    { domain "Research" trigger "Third-party API investigation, library evaluation, integration research" }
  ]
}

agent loom {
  prompt_append "Before implementing any third-party integration, always delegate to API Researcher first to understand the API surface."
}

Combining Mechanisms

The most powerful configurations combine multiple mechanisms:

# Optimize built-in agents
agent loom {
  models ["anthropic/claude-opus-4"]
  prompt_append "Prioritize security review for any auth-related changes."
}

agent thread {
  models ["anthropic/claude-haiku-4"]
}

# Add custom specialists
agent perf-analyzer {
  description "Performance Analyst"
  prompt "Analyze code for performance issues: N+1 queries, unnecessary re-renders, memory leaks, expensive computations in hot paths."
  models ["anthropic/claude-sonnet-4-5"]
  mode subagent

  tool_policy {
    read allow
    write deny
    execute deny
    delegate deny
  }

  triggers [
    { domain "Performance" trigger "Performance analysis, bottleneck identification, optimization review" }
  ]
}

# Disable what you don't need
disable agents ["spindle"]
disable hooks ["context-window-monitor"]

# Domain-specific categories for Shuttle
category backend {
  models ["anthropic/claude-opus-4"]
  prompt_append "Use async/await. Include comprehensive error handling."
}

Start Simple

Don't try to configure everything at once. Start with the default workflow, identify friction points, and customize incrementally. The most effective configurations are targeted: solving specific problems rather than trying to optimize everything.