Lobster Coordinator is a skill that implements a three-layer architecture for orchestrating multi-agent AI workflows. The three layers are AgentTool, which handles task delegation to individual agents; Swarm, which manages team-level messaging between agents; and Coordinator, which handles global task decomposition and result aggregation across the entire system.
The skill addresses a specific failure mode in multi-agent systems: agent loops, where agents get stuck in repetitive cycles without making progress. It tackles this through secondary verification layers and strict multi-stage review processes drawn from the Santa Method and Superpowers frameworks.
Code review follows a two-stage process. The first stage checks for specification compliance; the second checks code quality. The Santa Method adds a dual-blind verification step requiring two independent reviewers to approve code before it is merged. This reduces the risk of hallucinated outputs and accumulated technical debt in long-running agentic pipelines.
Model routing allows tasks to be assigned to different model tiers — Cheap, Standard, or Elite — based on their complexity, which can reduce inference costs. Agents report status as DONE, BLOCKED, or NEED_CONTEXT, giving the coordinator structured state information to manage task progression.
This skill is built specifically for advanced AI environments such as OpenClaw, Claude Code, and Cursor, and relies on capabilities like sessions_spawn for isolated sub-agent execution. It is not a general-purpose MCP server and has no database or API integration tools.