DenchClaw Subagents: Parallel AI for Complex CRM Tasks

DenchClaw's subagent system lets the AI delegate complex, time-consuming tasks to isolated parallel workers while keeping your main conversation responsive. Instead of waiting minutes for a bulk data enrichment or a browser automation task to complete, the main agent spawns a subagent, reports back to you immediately, and the subagent works in the background — announcing its results when it's done.

What Is a Subagent?#

A subagent is an isolated AI session spawned by the main agent to handle a specific task. Key properties:

• Isolated context: The subagent has no access to your MEMORY.md or main session history — it only knows what the main agent told it in its task brief
• Parallel execution: Multiple subagents can run simultaneously, each handling a different piece of a complex task
• Push-based completion: When a subagent finishes, it automatically announces its results back to the main agent
• Self-contained: A subagent does its assigned job and terminates — it doesn't persist

Think of subagents as the AI equivalent of spinning up background worker processes for CPU-intensive jobs.

When the Main Agent Uses Subagents#

DenchClaw's main agent automatically delegates to subagents for:

1. Long-running tasks that would block the main conversation (web scraping, bulk imports, browser automation)
2. Parallel workloads where multiple independent tasks can run simultaneously (researching 10 companies at once)
3. Specialist tasks requiring deep focus (building a Dench App, running a complex browser workflow)
4. Tasks with no user interaction required (data enrichment, bulk email drafting)

The main agent uses this heuristic: if a task takes more than ~3 steps or more than a few seconds, delegate to a subagent.

Example: Bulk Lead Enrichment#

Without subagents, enriching 50 leads would require the main agent to process them sequentially — potentially taking 10-15 minutes while blocking your conversation.

With subagents:

1. You ask: "Enrich all my San Francisco leads with company data from Apollo"
2. The main agent identifies 50 leads matching the filter
3. It spawns 5 subagents, each handling 10 leads in parallel
4. Main agent responds immediately: "I've spawned 5 enrichment workers. They'll process your 50 SF leads in parallel and report back when done."
5. You continue chatting normally
6. Within 3-4 minutes, the subagents complete and announce: "Enrichment complete: 47 of 50 leads updated with company size, industry, and contact email. 3 leads not found in Apollo."

The same task takes 4 minutes instead of 15, and you're never blocked.

Spawning Subagents Directly#

You can explicitly request parallel subagent work:

"Research these 5 competitors simultaneously: HubSpot, Salesforce, Attio, Folk, and Affinity. Find their pricing, key features, and recent funding news. Report back with a comparison table."

The main agent spawns 5 subagents, one per competitor. Each browses the web independently using the browser agent. All 5 complete within 2-3 minutes, and the main agent synthesizes their results into the comparison table you requested.

Without parallel subagents, this would require 5 sequential browser sessions.

Subagent Architecture#

Technically, each subagent is an OpenClaw session with:

• A unique session ID
• An isolated message history
• Access to all the same tools as the main agent (DuckDB, browser, filesystem, API calls)
• A task brief provided by the main agent at spawn time
• A one-shot execution model (it completes the task and terminates)

The main agent tracks which subagents it has spawned and waits for their completion events before synthesizing results.

Monitoring Subagents#

In the DenchClaw UI, active subagents appear in the status bar:

• A spinner icon with "3 agents running"
• Click to see each subagent's label, progress, and elapsed time
• Cancel a specific subagent if it's taking too long or went in the wrong direction

From the AI chat: "What are the current subagents working on?" — the main agent reports active tasks and their status.

Subagents for App Building#

Building a complex Dench App is a multi-step task that benefits from subagent isolation:

"Build me a dashboard app that shows my pipeline velocity by stage for the current quarter."

The main agent spawns an App Builder subagent with:

1. Your CRM schema (so it knows your field names)
2. The app brief (what it needs to show)
3. The app builder skill instructions

The App Builder subagent spends 2-5 minutes designing and building the app — writing HTML, CSS, JavaScript, and querying your DuckDB schema. Meanwhile, you're free to chat with the main agent, look at your CRM data, or work on something else.

When the app is ready, the subagent announces: "Pipeline Velocity Dashboard is ready. Find it in your Apps sidebar."

Subagent Depth Limits#

DenchClaw enforces a subagent depth limit of 2 (subagents can spawn sub-subagents for complex orchestration, but not deeper). This prevents runaway recursive delegation and keeps resource consumption bounded.

Resource Usage#

Each running subagent uses:

• Memory for its isolated session context
• API tokens for AI calls
• CPU/IO for any local operations

For large parallel workloads (many subagents running simultaneously), DenchClaw throttles spawning to avoid overwhelming your machine or hitting API rate limits.

See also: DenchClaw Cron Scheduling for scheduled subagent work, and DenchClaw Natural Language Queries for how the main agent processes your requests.

Frequently Asked Questions#

Can I spawn subagents manually, or does the main agent decide?#

Both. The main agent automatically delegates when appropriate. You can also explicitly request parallel work: "Run these 3 tasks in parallel". For advanced use cases, you can interact with the subagent system directly via the DenchClaw API.

Do subagents have access to my MEMORY.md?#

No. Subagents are isolated and don't load personal context files. This is by design — it keeps subagents focused on their specific task without the cognitive overhead of personal context, and prevents potential context leakage in sensitive data scenarios.

Can a subagent spawn more subagents?#

Yes, up to the depth limit of 2. A subagent handling a complex task (e.g., enriching 100 companies) may spawn its own subagents to parallelize. The main agent sees all of these as part of its orchestration tree.

What happens if a subagent fails?#

The main agent receives the failure announcement and reports the error to you. You can ask the main agent to retry the failed task, adjust the approach, or handle the failure gracefully (e.g., skip failed records and proceed with successful ones).

Are subagents billed separately?#

Subagents use the same AI model and API key as the main agent. Each subagent generates its own token usage, which counts against your API usage. For large parallel workloads, monitor your API token consumption. DenchClaw Cloud (dench.com) includes managed token limits in the subscription pricing.

Ready to try DenchClaw? Install in one command: npx denchclaw. Full setup guide →