Agent Audits
A capability audit for our builtin agents across scope, write access, delegation, and operational guardrails.
This is an internal posture view, not a third-party certification. It is meant to show where each agent sits on the spectrum from contained specialist to elevated orchestrator.
What we review
Four simple dimensions keep the audit readable and make tradeoffs visible at a glance.
Scope discipline
Each agent is reviewed against its intended role so broad access is reserved for agents that truly need it.
Tool boundaries
Read-only specialists stay read-only, while writing agents are limited to the execution paths they are built for.
Delegation surface
Only orchestration-focused agents can fan out work, which keeps parallel execution contained to explicit coordinators.
Verification posture
Execution agents are expected to verify with diagnostics, checks, or explicit review steps before calling work complete.
Current posture by agent
Higher control levels indicate broader execution authority or coordination surface.
| Agent | Scope | Access | Delegates | Control | Notes |
|---|---|---|---|---|---|
Sisyphus-Junior Focused task executor that completes delegated atomic tasks without spawning other agents. | Execution | Local writes | No | Guarded | Atomic task executor with explicit verification before handoff. |
Atlas Master orchestrator that drives a full todo list to completion by dispatching tasks to specialized agents in parallel waves. | Orchestrator | Local reads + delegation | Yes | Elevated | Highest coordination surface; intended for structured multi-wave work only. |
Prometheus Strategic planning consultant that interviews you to clarify requirements, then generates a structured work plan. | Planning | Plan output | No | Guarded | Writes work plans, but does not implement code changes directly. |
Oracle Read-only high-IQ reasoning specialist for architecture decisions, hard debugging, and post-implementation review. | Advisory | Read-only | No | Contained | Strategic reviewer with no write path and no task execution. |
Metis Pre-planning consultant that classifies intent, surfaces hidden requirements, and produces guardrail directives. | Pre-planning | Read-only | No | Contained | Clarifies intent and guardrails before implementation begins. |
Momus Work plan reviewer that checks plans for executability and valid references. Only blocks on true blockers. | Plan review | Read-only | No | Contained | Reviews plan executability and blocks only on concrete issues. |
Explore Contextual grep specialist for your own codebase. Fire multiple instances in parallel to find patterns and structure. | Repository search | Read-only | No | Contained | Finds code paths and structure without making edits. |
Librarian External reference specialist that finds official documentation, open-source implementations, and best practices. | External research | Read-only + docs | No | Contained | Pulls official references to reduce implementation guesswork. |
Hephaestus Autonomous deep worker that explores thoroughly before writing, then completes complex multi-file implementations end-to-end. | Autonomous execution | Local writes + external reads | No | Elevated | Broadest single-agent implementation surface; best for end-to-end work. |
Multimodal Looker Media file interpreter that extracts specific information from PDFs, images, and diagrams. | Media analysis | Read-only files | No | Contained | Extracts structured insight from images, PDFs, and diagrams only. |
Want to install one of these agents?
Browse the builtin catalog and copy the install command directly from the home page.