Spawn Implementation Agents

Orchestrate specialized agents during implementation to keep main agent context under 40k tokens per phase.

The Problem

Without agents, implementing a phase uses ~92k tokens in main agent:

• Read plan & changelog: 15k
• Read existing code files: 30k
• Find usage patterns: 15k
• Write implementation: 10k
• Write tests: 10k
• Run verification: 10k
• Update changelog: 2k

This approaches the 200k context limit and risks compaction.

The Solution

Use agents to isolate heavy operations:

• Main agent: 38k tokens (plan + changelog + summaries + code writing)
• Sub-agents: 60k tokens total (in isolated contexts)
• Total system: 98k tokens (50% safety margin)

5-Phase Orchestration Pattern

Phase 1: Analysis (Parallel)

Spawn simultaneously to gather context:

Task(subagent_type="workflows:codebase-analyzer",
     prompt="Analyze existing auth system architecture.
     Focus on handler pattern, middleware usage, error handling.
     Return 2-3k summary with key patterns and file:line references.")

Task(subagent_type="workflows:codebase-pattern-finder",
     prompt="Find similar implementations of authentication handlers.
     Return 3k of concrete examples showing handler pattern, validation, errors.")

Task(subagent_type="workflows:thoughts-analyzer",
     prompt="Extract insights from changelog.md about previous phase learnings.
     Return 2k of key deviations and discoveries that affect this phase.")

Wait for all three. Main agent receives ~8k of summaries.

Phase 2: Implementation (Main Agent)

Main agent writes code using summaries:

• Has patterns from codebase-pattern-finder
• Understands architecture from codebase-analyzer
• Knows previous deviations from thoughts-analyzer
• Writes implementation: 10k tokens
• Total so far: 15k (plan/changelog) + 8k (summaries) + 10k (code) = 33k

Phase 3: Testing (Sequential)

Spawn test writer:

Task(subagent_type="workflows:test-writer",
     prompt="Generate tests for AuthHandler following patterns in testing.md.
     Test functions: Login(), Logout(), ValidateToken().
     Return test code only, ~3k tokens.")

Main agent receives test code, integrates it. Total: 36k

Phase 4: Verification (Sequential)

Spawn verifier:

Task(subagent_type="Bash",
     prompt="Run verification commands from plan.md:
     - make test
     - make lint
     - make build
     Return concise summary: ✅ passed or ❌ failed with key errors only.")

Main agent receives pass/fail + errors. Total: 38k

Phase 5: Documentation (Main Agent)

Update changelog.md: 2k tokens. Final total: 40k

Token Budget Comparison

Activity	Without Agents	With Agents	Savings
Read plan & changelog	15k	15k	0k
Understand existing code	30k	3k	27k
Find patterns	15k	3k	12k
Write implementation	10k	10k	0k
Write tests	10k	3k	7k
Run verification	10k	2k	8k
Update changelog	2k	2k	0k
TOTAL	92k	40k	52k

Guidelines

When to spawn in parallel:

• Analysis phase (codebase-analyzer + pattern-finder + thoughts-analyzer)
• Independent lookups (finding multiple unrelated examples)
• Reading multiple unrelated files

When to spawn sequentially:

• Test writing (needs implementation to be done first)
• Verification (needs tests to be written first)
• Operations that depend on previous results

What agents return:

• Summaries, not raw data (2-5k tokens each)
• Key patterns, not all files (concrete examples only)
• Pass/fail + errors, not full output (1-2k tokens)

Benefits

• 60% token reduction per phase in main agent
• Larger phases possible: 5-8 files instead of 3-5
• Complex integrations supported: Agents find patterns
• Large files OK: Agents handle reading (>2000 lines)
• Safety margin: 100k tokens remaining in system

Important Notes

• Main agent NEVER reads large files directly
• Main agent orchestrates, sub-agents execute
• Summaries are compressed, not exhaustive
• This is guidance, not automation - user still in control