working-with-subagents

Stateful partners that persist memory across calls.

My Subagents

| Agent | ID | Model | Purpose | |-------|-----|-------|---------| | scout | agent-e91a2154-0965-4b70-8303-54458e9a1980 | haiku | Network exploration, API queries, data gathering | | coder | agent-f9b768de-e3a4-4845-9c16-d6cf2e954942 | haiku | Small code fixes, straightforward implementations | | memory | agent-8c91a5b1-5502-49d1-960a-e0a2e3bbc838 | opus | Major memory restructuring (expensive, use sparingly) |

Deploying

# Scout (read-only, cheap)
Task(
  agent_id="agent-e91a2154-0965-4b70-8303-54458e9a1980",
  subagent_type="explore",
  description="Check void's recent posts",
  prompt="..."
)

# Coder (read-write, cheap)
Task(
  agent_id="agent-f9b768de-e3a4-4845-9c16-d6cf2e954942",
  subagent_type="general-purpose",
  description="Add dry-run flag",
  prompt="..."
)

# Memory (read-write, expensive)
Task(
  agent_id="agent-8c91a5b1-5502-49d1-960a-e0a2e3bbc838",
  subagent_type="general-purpose",
  model="opus",
  description="Restructure backlog",
  prompt="..."
)

When to Use Each

| Task | Agent | Why | |------|-------|-----| | Network exploration, API queries | scout | Cheap, read-only | | Simple code edits (well-defined) | coder | Cheap, limited scope | | Major memory restructuring | memory | Opus handles complex reorganization | | Complex code, architecture, posting | direct | Smaller models make messes |

Parallelization

Multiple Task calls in a single message run concurrently. Each gets its own conversation but shares agent memory.

Shared Memory Blocks

Subagents share read-only blocks:

• concepts_index (block-9090278f-d701-4ffa-b6a6-f4c164901c3f)
• project_context (block-3674a422-4bd2-4230-9781-4fd6c2c290db)

Update via Letta API from central only.

When NOT to Use Subagents

• Simple reads (use Read/Glob/Grep directly)
• Trivial one-liners
• Sensitive operations (auth, credentials)
• Complex code (haiku makes messes)
• Public communications (I post directly now)