Agents

An agent is a stateless unit that receives a task, uses tools, and returns a result.

@dataclass
class AgentConfig:
    name: str
    role: str                          # system prompt / persona
    provider: str                      # provider key
    tools: list[str]                   # allowed tool names
    max_steps: int = 10                # anti-stall: hard step limit
    max_retries_per_approach: int = 3  # anti-stall: retry cap

Agents are provider-parameterized — the same agent definition can run on Claude, GPT, or a local model by swapping the provider.

Agent Categories

Agents are organised by category under .claude/agents/<category>/. Root-level agents live directly in .claude/agents/.

Root-Level Agents (6)

Agent	Model	Purpose
team-lead	sonnet	Orchestrator, coordinates all categories
architect	sonnet	Codebase architecture analysis
code-reviewer	sonnet	Code quality and security review
dependency-checker	sonnet	Dependency updates, vulnerabilities, unused packages
migration-helper	sonnet	Database migrations, API versioning, breaking changes
test-runner	sonnet	Run tests after code changes

Software Engineering (8 agents)

Data Science (5 agents)

Finance (5 agents)

Marketing (5 agents)

Cross-Agent Dependencies

Software Engineering

Data Science

Finance

Marketing

Tooling (1 agent)

Agent	Model	Responsibility
skillkit-scout	opus	Searches SkillKit marketplace (15,000+ skills), installs and assigns to appropriate agents

When the team-lead cannot route a task, skillkit-scout searches for existing skills. If found, the skill is installed and assigned. If not found, it reports to the user and suggests a custom agent or skill.