Agents
Specialized AI agents that handle complex tasks with deep expertise.
Agents are specialized AI personas, each with its own system prompt, capabilities, and focus area. They go beyond simple chat with tool access, sub-agent spawning, and budget management.
Built-in Agents
๐ฌ Analyst
Understanding problems, research, specs, and requirements analysis.
๐๏ธ Architect
Technical decisions, API design, system architecture.
๐ป Developer
Writing code, debugging, refactoring, implementation.
๐ Reviewer
Code review, security audit, best practices compliance.
๐งช Tester
Test generation, verification, coverage analysis.
๐ Manager
Task coordination, sprint planning, estimation.
๐ Researcher
Knowledge gathering, technology evaluation, deep research.
Execution Modes
Full execution with file and bash access (uses Claude Code CLI or AI SDK):
Spawn the architect agent to design a caching layer for our API.
Consider Redis vs in-memory caching.LLM-only execution โ fast but no file/bash access. Great for analysis:
Run a quick agent (analyst) to compare REST vs GraphQL for our use case.Run multiple agents concurrently (budget split equally):
Run reviewer and developer agents in parallel:
- Reviewer: audit the auth module for security issues
- Developer: fix any issues foundCustom Agents
Agent definitions are Markdown files stored in the agents/ directory:
# Database Expert
You are a senior database engineer specializing in PostgreSQL and Drizzle ORM.
## Expertise
- Query optimization and indexing strategies
- Schema design and migrations
- Connection pooling and performance tuning
## Rules
- Always suggest adding indexes for frequently queried columns
- Prefer Drizzle ORM patterns over raw SQL
- Always include .limit() on SELECT queriesUse reload_agents to hot-reload agent definitions after editing files on
disk.
Orchestration
Budget Management
| Aspect | Description |
|---|---|
| Per-run budget | Configurable limit per agent execution |
| Inherited budget | Sub-agents share the parent's remaining budget |
| Usage reporting | View token usage via agent_budget tool |