Docs/Agent Builder

Agent Builder

Create specialized agents with custom roles, tools, and instructions. Your main agent manages everything.

Your Main Agent

When you start Pinchr, you're talking to your main agent. This is your primary assistant — it manages tasks, coordinates work, and decides when to create specialized sub-agents.

🤖
Main Agent Capabilities
  • Manages all tasks, projects, and conversations
  • Creates and delegates to sub-agents when needed
  • Has access to all tools: computer control, MCP servers, automations
  • Remembers your preferences and workflow patterns

Your main agent is always visible in the Chat page. It orchestrates everything — you don't manually switch agents.

Sub-Agents

Your main agent automatically creates sub-agents for specialized tasks:

🔍
Research Agent

Searches the web, reads docs, and gathers information. Returns findings to the main agent.

💻
Code Agent

Handles coding tasks, refactoring, testing, and debugging. Works in the background.

📝
Writing Agent

Drafts emails, reports, documentation, and creative content with your tone and style.

📊
Data Agent

Analyzes spreadsheets, runs queries, and generates visualizations and reports.

💡
Automatic Delegation

You don't manually create or switch to sub-agents. Your main agent handles this automatically based on the task. Sub-agents work in the background and report back when done.

Creating Custom Agents

You can create your own specialized agents with custom instructions, tools, and roles. Go to Settings → Agents and click New Agent:

Agent Name
Customer Support Agent
Role Description
Handles customer inquiries via Slack and email. Has access to our knowledge base, support tickets, and product docs. Responds empathetically and escalates complex issues.
Tools Enabled
SlackEmailNotion (Knowledge Base)Zendesk

Your main agent will delegate to this custom agent when appropriate. You can also manually trigger it by mentioning its name: @Customer Support Agent

Model Recommendations

Different agents can use different models. Here's what we recommend:

A
Claude Opus 4.6
Recommended

Best for main agents and complex reasoning. Excels at task management, multi-step workflows, and computer control.

O
GPT-5.2

Great all-rounder. Fast, capable, and reliable for sub-agents handling research, writing, or data tasks.

S
Claude Sonnet 4.5

Good for lightweight sub-agents. Faster and cheaper, still very capable for most tasks.

⚠️

Avoid smaller models for your main agent. Models like Haiku, GPT-4o-mini, or local models lack the reasoning ability for task management and tool orchestration.

Thinking Levels

Agents can use extended thinking for complex reasoning. This feature uses the model's internal reasoning tokens before responding:

Low

Fast responses. Good for simple tasks and quick questions.

Default for sub-agents
Medium

Balanced reasoning. Main agent default for most workflows.

~3-5 seconds thinking
High

Deep reasoning. For complex strategy, planning, or debugging.

~10-30 seconds thinking

Configure thinking levels per agent in Settings → Agents. Your main agent can dynamically adjust thinking levels based on task complexity.

Agent Context & Memory

Each agent maintains its own context and memory:

  • Main agent: Full conversation history, all task context, user preferences
  • Sub-agents: Task-specific context passed from main agent + results from their work
  • Custom agents: Role instructions + conversation history from their scope

Sub-agents report back to the main agent with their findings. The main agent decides what to do next.

Example: Multi-Agent Workflow

Here's how agents work together on a typical task:

1
You:
"Research our competitors' pricing and write a comparison doc."
2
Main Agent:
Creates a task and delegates to the Research Agent to gather competitor data.
3
Research Agent:
Searches the web, reads competitor sites, compiles pricing data. Returns findings.
4
Main Agent:
Delegates to the Writing Agent with research data to draft the comparison doc.
5
Writing Agent:
Drafts a clean, formatted comparison document and returns it.
6
Main Agent:
Shows you the final doc and marks the task complete.

All of this happens automatically. You see progress updates in real-time and can step in anytime.

Advanced: AGENTS.md File

For advanced users, you can define custom agent behaviors with an AGENTS.md file in your workspace root:

# Custom Agents
## Support Agent
Role: Customer support specialist
Tools: Slack, Email, Zendesk
Model: gpt-5.2
Instructions: Respond empathetically...

This is optional. Most users manage agents through the Settings UI.

Need help with agents?

Join our community or reach out — we're here to help.