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)
1---2name: working-with-subagents3description: Guide for deploying and prompting my stateful subagents (scout, coder, memory). Use when delegating tasks or parallelizing work.4---56# Working with Subagents78Stateful partners that persist memory across calls.910## My Subagents1112| Agent | ID | Model | Purpose |13|-------|-----|-------|---------|14| **scout** | `agent-e91a2154-0965-4b70-8303-54458e9a1980` | haiku | Network exploration, API queries, data gathering |15| **coder** | `agent-f9b768de-e3a4-4845-9c16-d6cf2e954942` | haiku | Small code fixes, straightforward implementations |16| **memory** | `agent-8c91a5b1-5502-49d1-960a-e0a2e3bbc838` | opus | Major memory restructuring (expensive, use sparingly) |1718## Deploying1920```python21# Scout (read-only, cheap)22Task(23 agent_id="agent-e91a2154-0965-4b70-8303-54458e9a1980",24 subagent_type="explore",25 description="Check void's recent posts",26 prompt="..."27)2829# Coder (read-write, cheap)30Task(31 agent_id="agent-f9b768de-e3a4-4845-9c16-d6cf2e954942",32 subagent_type="general-purpose",33 description="Add dry-run flag",34 prompt="..."35)3637# Memory (read-write, expensive)38Task(39 agent_id="agent-8c91a5b1-5502-49d1-960a-e0a2e3bbc838",40 subagent_type="general-purpose",41 model="opus",42 description="Restructure backlog",43 prompt="..."44)45```4647## When to Use Each4849| Task | Agent | Why |50|------|-------|-----|51| Network exploration, API queries | **scout** | Cheap, read-only |52| Simple code edits (well-defined) | **coder** | Cheap, limited scope |53| Major memory restructuring | **memory** | Opus handles complex reorganization |54| Complex code, architecture, posting | **direct** | Smaller models make messes |5556## Parallelization5758Multiple Task calls in a single message run concurrently. Each gets its own conversation but shares agent memory.5960## Shared Memory Blocks6162Subagents share read-only blocks:63- `concepts_index` (block-9090278f-d701-4ffa-b6a6-f4c164901c3f)64- `project_context` (block-3674a422-4bd2-4230-9781-4fd6c2c290db)6566Update via Letta API from central only.6768## When NOT to Use Subagents6970- Simple reads (use Read/Glob/Grep directly)71- Trivial one-liners72- Sensitive operations (auth, credentials)73- Complex code (haiku makes messes)74- Public communications (I post directly now)