Beads Integration Patterns
Issue Tracking for AI Agent Workflows
Abstract
How can issue tracking systems be designed to support nondeterministic, context-limited AI agent workflows? This paper analyzes Beads integration patterns across the CREATE SOMETHING monorepo, documenting patterns for context survival, work extraction, and dependency tracking that enable agents to pick up where they left off.
Research Question
Traditional issue trackers assume human operators with continuous context. AI agents face different constraints: context windows fill, sessions restart, and multiple agents may work in parallel. How do we design persistence that survives these conditions?
Methodology
Analyzed Beads implementation across multiple packages:
/packages/harness/src/beads.ts/packages/agent-sdk/src/create_something_agents/tools/beads.py/.claude/rules/beads-patterns.md
Integration Patterns
Context Survival
Preserve work across session restarts and context limits
bd create "Refactor authentication"
bd update cs-123 --status in-progressWork Extraction
Convert review findings into actionable issues
createIssueFromFinding({
severity: 'high',
title: 'Add input validation',
description: 'Prevent potential security risks'
})Dependency Tracking
Model complex work relationships and blockers
bd dep add cs-auth blocks cs-dashboardWork Discovery Taxonomy
Issues emerge from three sources during agent execution:
blocker
Critical findings that immediately halt progress
Critical security vulnerability requiring immediate fixsupervisor
Standard review findings about architecture, quality
Performance optimization suggestionrelated
Secondary work items discovered during primary task
Potential refactoring during feature implementationLimitations
- Relies on CLI tool availability across environments
- Potential race conditions in multi-agent environments
- Requires consistent tooling across development environments
- Git-based persistence adds commit overhead