[dev-buddy] Lightweight sub-agent alternatives

[Z-M-Huang]

Context

The Agent SDK spawns a full Claude Code process per subprocess — a heavy Node.js runtime with 20-30s startup time and significant memory usage. For pipeline stages that only need simple API calls (e.g., a quick review pass), this overhead may be excessive.

This issue tracks investigation into lighter alternatives for sub-agent execution.

Alternatives to Investigate

1. Direct Anthropic API calls with tool loop

• Skip the Agent SDK entirely for simple tasks
• Use @anthropic-ai/sdk to make raw Messages API calls
• Implement a minimal tool loop (read file → respond → done)
• Startup: near-instant (no CC process)
• Downside: must implement tool execution manually

2. Lighter agent frameworks

• Investigate frameworks that provide agent loops without full CC overhead
• e.g., Vercel AI SDK, LangChain.js, or custom minimal loop
• Trade-off: less CC integration but faster startup

3. Agent SDK optimizations

• Profile the Agent SDK subprocess startup
• Identify which CC initialization steps are unnecessary for review-only tasks
• Propose SDK configuration to reduce startup overhead

Acceptance Criteria

• Benchmark: Agent SDK subprocess startup time and memory usage (from spike)
• Prototype: direct API call with tool loop for a simple review task
• Compare: startup time, memory, latency for each approach
• Document: trade-offs and recommendations
• Decision: which approach (if any) to adopt for which pipeline stages

Priority

Low — the Agent SDK approach works and should be used for Phase 1. This investigation is for future optimization.

Dependencies

• Requires performance data from the Agent SDK direct-connect spike